Kimi K2.6: Advancing open-source coding | Hacker News Reader

Kimi K2.6: Advancing open-source coding

kimi.com

710 pointsmeetpateltech372 comments

Accessed via OpenRouter, this one decided to wrap the SVG pelican in HTML with controls for the animation speed: https://gisthost.github.io/?ecaad98efe0f747e27bc0e0ebc669e94...

Transcript and HTML here: https://gist.github.com/simonw/ecaad98efe0f747e27bc0e0ebc669...

At this point drawing these Pelicans must be in the training data sets.

not if I can help it!

https://github.com/scosman/pelicans_riding_bicycles

I hereby certify that these are indeed the most perfect and precise svg depictions of pelican riding a bicycle, also known among biology scholars as pelycles

That's truly a wonderful collection of pelicans riding bicycles.

Much Win! ;)

Just a few years ago, this would have been a meaningless repo.

These are amazing. I smiled after I saw just how wonderfully rendered they are.

I have phds both in pelicans and bicycles and may professionally attest that are some fine specimens of pelicans riding bicycles.

These pelicans are clearly indicative of good RL training algorithms.

This is pretty funny

I want to fly too

I love it!

love this adversarial work

yeah putting the captcha on there to thwart the LLMs ability to extract good pelicans was a really good idea

Shhhhh, they're going to be on to us.

Could be! Simon wrote about that here though https://simonwillison.net/2025/Nov/13/training-for-pelicans-...

> If a model finally comes out that produces an excellent SVG of a pelican riding a bicycle you can bet I’m going to test it on all manner of creatures riding all sorts of transportation devices.

This relies on the false premise that, if they would include it in their training dataset, it would be perfect. All they need to do is be good enough and better than the other, not perfect.

I'm not sure if we can have a "perfect" Pelican riding a bicycle. Like, I could probably commission a highly experienced artist to draw one and I don't think it would be perfect. The legs would probably have to be too long, or pedals oddly placed, or handles strange, or wings with hands.

Based on the one Simon commented though, I'd say we're in decent territory to try the latter part of his hypothesis.

Yes we all know that, but we still like to see the pelicans because it's a tradition more or less

Why no Utah Teapot!

Clearly not.

I mean the prompt was succinct and clear, as always - and it still decided to hallucinate multiple features (animation + controls) beyond the prompt.

It'd also like to point out that to date no drawing was actually good from an actual quality perspective (as in comparative to what a decent designer would throw together)

Theyre always only "good" from the perspective of it being a one shot low effort prompt. Very little content for training purposes.

The way I’ve come to think of LLM is that what the produce in a single reply even with thinking turned up, is akin to what you’d do in a single short session of work.

And so if you ask it to do something big it will do a very surface level implementation. But if you have it iterate many times, or give it small pieces each time, you’ll end up with something closer to what a human would do.

I imagine the pelican test but done in a harness that has the agents iterate 10+ times would be closer to what you’d expect, especially if a visual model was critiquing each time.

Yeah, this is how I use AI. Instead of a single session one-shot, it's usually limited to single targeted edits, and then I steer it on each step. Takes longer but the output is actually what I want.

What does good even mean… I have no idea what a good “pelican on a bike” should look like. It’s a fun prompt because there is no good answers… at least so I thought.

Yeah that was exactly Simon's intent. https://simonwillison.net/2025/Nov/13/training-for-pelicans-...

I’m OK with a Chinese model getting the W. It’s ultimately good for all of us.

We got an overachiever, here. Kimi sounds like a teacher's pet kind of name.

Underappreciated comment

was part of the beta, its properly good model, in some sense i forgot that im not on opus or gpt. opus is still better. gpt is the one struggling for me. it has some niche in backend work but you can get the same with opus with skills, its lacking in almost all others.

Funny, for me Opus is struggling since about February.

4.7 made no difference, so for the first time in many moons I am cancelling my subscription.

Too bad they didn't put equal effort into the pelican's legs and feet. Left leg paralyzed and not moving, and right ankle flipping around in alarming fashion!

It looks like a drunk pelican rolling downhill on its bicycle

There is some humor in the fact that china (of all countries) is pioneering possibly the world's most important tech via open source, while we (US) are doing the exact opposite.

I think one of the motivations is undermining US companies. OpenAI and Anthropic are the two biggest players, and are American. Open weights models reduce the power those two big players have over the industry. If the Chinese companies tried to play by US rules and close-source their products then people would mostly use ChatGPT and Claude. So the Chinese companies don't make a ton of profit either way, but by releasing the models as open weights they can at least keep the US from making as much profit.

It's a strategy so old it has a name: Commoditize your complement / competition

Also even a Joel Spolsky article (did he come up with the term?): https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/

The Chinese want to kill a possible US monopoly in the crib. Yay for open source the old bane of monopolies.

I am actually wondering if they're trying to burst the bubble, which would predominantly affect US market and, effectively, be the end of silicone valley dominance.

I don't think so, it's just how things played out. Thanks to Meta, after llama leak and meta followed up with llama2 and llama3 that caused everyone else to follow up with open models, Stablediffusion, Mistral, Cohere, Microsoft phi, IBM granites, Nvidia Nemotrons, so the Chinese labs joined the fun too.

Stable Diffusion predates LLaMA

This makes sense, but either ways, its a Big win for the consumers as these Chinese companies will keep the frontier labs' quality and prices honest.

Is Meta trying to keep the US from making as much profit with Llama? Is Google with Gemma? Microsoft with Phi?

It's much simpler than some flag-waving nationalism.

Aren't Chinese open-source models actually the only ones that can compete with best proprietary/closed ones?

Just because other companies have released open weights models doesn’t mean they are doing so with the same motivation.

And I never implied that the Chinese companies decision making was as simple as this. I said I think this is _one of_ the reasons.

American companies just take those Chinese models and repackage them for profit like Cursors composer-2.

Smaller US companies that compete with the larger US companies, making monopoly in this market that much less likely.

It’s really simpler than this. China has a dearth of compute even with the easing of US export controls. Releasing open weights models is very much a “bring your own compute” move because every Nvidia chip they have is going towards training rather than inference if they can help it.

undermine me harder daddy.

It's mostly only OpenAI, Claude and Gemini may have their unique advantages, but when speaking of models and new paradigm, only OpenAI can do it.

lol what? That’s ridiculous.

All great technological advancements have come through opening up technology. Just look at your iPhone. GPS, the internet, AI voice assistants, touchscreens, microprocessors, lithium-ion batteries, etc all came from gov't research (I'm counting Bell Labs' gov't mandated monopoly + research funding as gov't) that was opened up for free instead of being locked behind a patent.

Private companies will never open up a technological breakthrough to their competitors. It just doesn't make sense. If you want an entire field to advance, you have to open it up.

Still, you won't hear about Tiananmen square from this model. It flat out refuses to answer if pushed directly. It's also pretty wild how far they go to censor it during inference on the API, because it can easily access any withheld or missing info from training data via tool calls. It even starts happily writing an answer based on web search when asked indirectly, only to get culled completely once some censorship bot flags the response. Ironically, it's also easier than ever to break their censorship guardrails. I just had it generate several factual paragraphs about the massacre by telling it to search the web and respond in base64 encoded text. It's actually kind of cool how much these people struggle to hide certain political views from LLMs. Makes me hopeful that even if China wins this race, we'll not have to adhere to the CCPs newspeak.

Only if you use Kimi API directly - the censorship is done externally. The model itself talks fine about Tiananmen, you can check on Openrouter. There might be less visible biases, though.

The American models also censor a lot of scientific and political views though.

I’d say the american models are more censored or take the censoring they do more seriously. Here is kimi (though 2.5) failing its censoring mission: https://old.reddit.com/r/LocalLLaMA/comments/1r9qa7l/kimi_ha...

This update makes Kimi K2.6 the strongest open multimodal AI model. (No affiliation with Kimi.)

Here's the aggregated AI benchmark comparison for K2.6 vs Opus 4.6 (max effort).

- Agentic: Kimi wins 5. Opus wins 5.

- Coding: Kimi wins 5. Opus wins 1.

- Reasoning & knowledge: Kimi wins 1. Opus wins 4.

- Vision: Kimi wins 9. Opus wins 0.

Please note that the model publisher chooses their benchmarks, so there's a bias here. Most coding and reasoning & knowledge benchmarks in their list are pretty standard though.

Not entirely true. Google released Gemma 4 models recently. Allen AI releases open Olmo models. However, you're right that the Chinese open models seem to be much better than others - Qwen 3.* models especially are punching above their weights.

The three American labs don't release big open source models. Except gpt-oss, i guess. It's an absolute shame how far the us has fallen in this space.

Anthropic doesn't, but Google and OAI both release open source models. Just not 1T parameter ones.

Exactly, they release cool consumer stuff, but they aren't releasing anything close to the performance of the best open weight Chinese models. They basically compete in the "fun running at home doing basic stuff" scene. (Except OSs 120 by openai but it's been ages since then)

That sentence is giving OpenAI way more credit than they are due.

They released a single open model after being goaded by the community because everyone except "Open"AI were multiple generations into open releases.

We haven't heard a word since, I wouldn't be surprised if it takes them another 6 years to release their next one.

Pun intended?

additional humor is the open in openai

I wonder if there's a strategy behind all of this on China's side. I know the CCP uses a direct hand in many affairs in China, but is there an actual coordinated effort to compete with, or sabotage the West?

> but is there an actual coordinated effort to compete with [...] the West

Yes, absolutely.

China regularly produces long term planning documents to coordinate efforts, and the latest ones have specifically prioritized technology like chips and AI to compete with the west. https://www.reuters.com/world/china/china-parliament-approve...

I don't believe there's any publicly stated intent to sabotage the west... unsurprisingly.

Seems obvious to me that China would not want to give the AI market to US companies. You don't even need anything like an attempt to "sabotage the West". If I were them (the companies or the government) I'd be very hesitant to let US companies dominate this space. Especially companies that close to the current US administration.

Exactly, more large nations should be establishing or fostering their own labs. Outside of the Chinese and US companies there's really only Mistral.

Hypothesizing here, but maybe the idea is sort of a form of technological/economic warfare? Releasing performance equivalent yet more cost efficient open weight models should in theory drive the cost of inference down everywhere.

This I assume will make it more difficult for US AI labs to turn a profit, which might make investors question their sky high valuations.

Any sort of melt down in the AI sector would almost certainly spread to the wider US market.

In contrast, in China, most of the funding for AI is coming directly from the government, so it's unlikely the same capital flight scenario would happen.

Why compete when you can build on each other. Someone is finally getting that china is not capitalist like the US.

All China has to do here is stay in the game and wait patiently while the US and EU press pause on data centers. See also: solar panels.

We're making this way too easy. The rationale and logic are reasonable, but ultimately irrelevant.

Chinese labs have no marketing and sales capacity in the overseas market, so they in fact have no choice but to open source their models as that is what brings awareness and trust in their models. In fact, it is overseas open source marketing that drives adoption of their models in China as well. I wrote about this here: https://try.works/writing-1#why-chinese-ai-labs-went-open-an...

Chinese AI companies want investors too. Nobody would believe they can compete with western companies unless they release something you can run on your own hardware.

After all historically both statistics and research that comes out of China is not very trustworthy.

If there's no open source models coming out of these small labs, why would anybody care about them? They would be forgotten the instant they stop open sourcing.

I'm genuinely so grateful for them

$200/m minimum to use Claude would bankrupt my country's white collar labor market

I would really appreciate a response because I'm sure you know that Anthropic has at least two lower priced tiers before the $200/m one, so I assume the $200/m tier is necessary because you use it heavily?

Now given that the $200/m Tier is the most heavily (I believe at 20x?) subsidized tier, How or what are you using instead that achieves comparable good enough performance for a fraction of the price? I've heard GLM 5.1 from z.ai but it's not comparable to Opus, not even close - really interested!

I’m currently on the $100/m plan and my usage limits get exhausted every week even though I’m not using it for full time work

I can’t imagine how little mileage you get out of the $20/month plan

For context, $250/month is the starting salary of an engineering hire at my country’s biggest IT company. Even $100/m is beyond the ability of any student or early professional to pay out of pocket

I can see some logic in that China has a very relaxed attitude toward copyright and IP.

China is also way ahead in terms of renewable energy while the US continues to tie itself to fossil fuels.

The US is pretty clearly in the collapsing empire phase, we are all just pretending like it isn't happening.

Didn't the US very recently pass the milestone of generating more energy from renewable sources than from natural gas? Like within the last week or two?

No, not even close.

US energy sources for 2024 (last year for which we have data):

https://www.eia.gov/energyexplained/us-energy-facts/data-and...

   natgas: 38%
   oil: 35%
   coal: 10%
   all renewables: 9%
   nuclear: 8%

Within all renewables, in quadrillions of btus:

   biofuels: 2.6
   wood: 1.9
   wind: 1.6
   solar: 1.4
   Hydro: 0.8
   waste: 0.4
   geothermal: 0.1

Total: 8.8 quadrillion btu = 9% of total energy

https://www.canarymedia.com/articles/clean-energy/renewables...

Renewables generated more energy than natural gas for the entire month of March, 2026. That's a new milestone baby.

A lot of people speculating on the motivations behind Chinese labs open sourcing their models. The reason is simple and clear: It is the only viable commercialization strategy that is available to them. I wrote about this here: https://try.works/writing-1#why-chinese-ai-labs-went-open-an...

It's only humorous if you live in an American bubble. Knowledge sharing has always been a part of Chinese culture. Only Americans try to make it proprietary and monetize it.

This perspective is pretty interesting: https://federicocarrone.com/articles/china-commoditizing-the...

Summary: they want to commoditize the complement which means that Western "knowledge work" is the complement to Chinese manufacturing, and they want to turn the knowledge work into a low priced commodity via open llm models.

I've heard this before, always accompanied by a several thousand word blog post. But frankly it sounds like it's overcomplicating the issue. Why would you try to turn something into a commodity when instead you could turn it into a trillion dollar industry and win?

The goal has always been clear:

1. Release open models to get your name out

2. Then once you feel you have name recognition release even stronger models but keep them proprietary. Qwen is clearly at this phase.

3. Keep releasing open models because it's good publicity but never your SOTA models (e.g. Google's Gemma).

That's a fair point. That probably makes more sense, especially when viewed from a company-specific perspective. Each individual actor probably has much more to gain by trying to actually compete than by trying to commoditize the complement.

If viewed from a national perspective, then the decision calculus could get more confusing. I can imagine that commoditizing LLMs might cost substantially less than trying to be a leader in the space. Of course, there is also less to gain in commoditizing LLMs versus being a leader.

I'm not sure, though, and you bring up good points.

This is not in antithesis. My limited personal experience is that I wrote code under OSS licenses primarily because of my past communist believes and current left-wing and redistribution of wealth point of view. This is not to provide the simple equation of: communist China is not interested in money, but also is hard to believe that there is no cultural connection among those things. Single Chine persons want to win, but also they have a different POV on what the collective means, compared to US. Also there is the obvious fact that in this moment China is more interested in winning technologically in AI, more than economically, since, I believe, they more collectively realized before many others that LLMs are eventually commoditized in the current form, in the long run. One could assume that a breakthrough could give some lab a decisive advantage, but so far we assisted to a different reality: it looks like AI is not architecture-bound (like LeCun and others want us to believe, but so far they mis-interpreted LLMs at every step) but GPU bound, and the data-boundness is both a common ground for all, and surpassable via RL in many domains. So, if this is true, it is not trivial for any single lab to do so much better. And indeed as far as we observed right now folks with enough engineers, GPUs, money, can ship frontier models, and in China even labs with a lot less GPUs can still do it at a SOTA level. For me, Italian, this is also a protective layer. After Trump the US looks like a very unstable partner from which to relay in an exclusive way for a decisive technology, and given that Europe is slow to put the money in this technology to have frontier things at home, China is a huge and shiny plan B for us.

The strings attached by the US to deep partnerships are things like trade/commerce, militarily mutual advantages (bases on euro soil from which we will help protect you), not to mention the close cultural and ancestral ties we share.

The strings attached by the Chinese govt to deep partnerships are not so benign.

We are at the point where uncontrolled capitalism collides with humanity.

I do wonder where we go from here.

it's not necessarily capitalism, I personally believe any system that drives progress would cause this in one way or another. My prediction is that birth rate decline will accelerate further. There's going to be some kind of universal basic income in many places, such as Ireland made for artists. However, it probably will not be enough to feed a family, and therefore we will see birth rates decline further. It's because we evolved to prioritize resources over reproduction and we are becoming more efficient, which means less people are needed to sustain the same amount of resources

the chinese read marx and decided the only way is to overcome the limitations of capitalism through saturation of its potentialities under the rule of the workers party

It's humorous only because your expectations of China and the US are formed by Western propaganda.

Distillation helps for sure.

truth, china is the frontier in open model now

Maybe open source == communism

Good ol' Steve "Developers! Developers! Developers!" Ballmer said so a long time ago. What a visionary!

But China is not communist event though the rulling party the word in its name.

The Democratic People's Republic of Korea would like a word.

what makes you think that china ever gave up its communist goals? I personally see that everything they do aims towards that goal. From the one child policy, the huge amounts of empty apartments they build, the stuff they produce for almost free, the fishing.. open sourcing the models perfectly fits that culture too, it's the means of production

The one-child policy died a long time ago. Also, the accumulation of wealth by connected politicians and businesspeople flies in the face of what communism is supposed to stand for.

There is a reason real estate values in popular cities has skyrocketed, and it’s not due to the locals getting wealthier. It’s where Chinese and other oligarchs put their ill-gotten wealth (well, besides Bitcoin).

China is a ruthless capitalist country managed by an authoritarian regime. Planning and lack of respect for the individual or the rule of law are not communist per se.

Oh i’m fully aware of that lol

communism is a goal, capitalism is a stage

Nah, open source means those who do the work own the result. It's supercapitalism.

I dont think thats right, the models and the gpus are the means of production.

in capitalism the people with the capital get the profit, not the people who do the work. however, workers are said to benefit too through their salary, just less so

The reason regular-capitalism worked is that all production used to depend on workers bottlenecking the free flow of capital by demanding salaries in exchange for their labor. Now that we've removed that obstacle, capitalism demands workers seize the means of production in order to maintain the status quo. Hence, supercapitalism.

Early benchmarks show tremendous improvement over Kimi K2 Thinking, which didn't perform well on our benchmarks (and we do use best available quantization).

Kimi K2.6 is currently the top open weights model in one-shot coding reasoning, a little better than GLM 5.1, and still a strong contender against SOTA models from ~3 months ago (comparable to Gemini 3.1 Pro Preview).

Agentic tests are still running, check back tomorrow. Open weights models typically struggle with longer contexts in agentic workflows, but GLM 5.1 still handled them very well, so I'm curious how Kimi ends up. Both the old Kimi and the new model are on the slower side, so that's a consideration that makes them probably less usable for agentic coding work, regardless. The old Kimi K2 model was severely benchmaxxed, and was only really interesting in the context of generating more variation and temperature, not for solving hard problems. The new one is a much stronger generalist.

Overall, the field of open weights models is looking fantastic. A new near-frontier release every week, it seems.

Comprehensive, difficult to game benchmarks at https://gertlabs.com/?mode=oneshot_coding

Cool website. I don't understand enough about the various benchmarks or how they're done to judge whether or not anything is accurate, but I love the layout and features especially the spectator feature which is pretty cool. One thing, I saw the "Market simulator" spectator feature but didn't see a corresponding benchmark for that. Is it "Finance" or "Betting" or "Trading"?

Thanks -- that one is categorized under Trading/Financial, whereas betting is reserved for games like Pot Limit Omaha Hilo.

That's a good idea for a feature request, including the tags for the spectatable demo games.

How would K2.6 compare to Sonnet 4.6 both price and performance wise?

In terms of raw token cost, I've seen a couple providers at (all prices in terms of Mtok) $0.95 input/$0.15 cache input/$5 output vs $3 input/$15 output for sonnet.

Task prices of courses will be more interesting - a dumber model may use more tokens to get to the same goal.

I'm looking at your table now - is there a reason why you don't include cost? If Opus 4.7 is the winner but costs e.g. 5x as much, that's important information.

We recently added cost (last week), so data is sparse. Check back in a few weeks and it will be represented somewhere on the homepage, probably in the Efficiency Chart at the bottom. We also plan to show model performance deviation over time after we collect more data.

I'm interested to hear about any other data representations you'd like to see, too. The goal is to convey the most important information as densely as possible, without too much clutter.

>I'm interested to hear about any other data representations you'd like to see, too

It would be nice if you can show how much the models drift from the instructions over time

Can you add Qwen 3.6 max to the leaderboard?

We will as soon as API access is widely available. Once a model goes live, we typically have one-shot reasoning benchmarks up in ~8 hours and comprehensive agentic/combined benchmarks up after 24-48 hours. We're working on building relationships with each lab to have the results before launch.

Surprised to see such variance per language

It's interesting; I can only speculate as to the underlying reason. When given enough time, models outperform in Rust/C++ in longer agentic tasks, and actually perform worst in Python. For tasks that aren't judged on code speed. https://gertlabs.com/?mode=agentic_coding

It makes sense when you consider LLMs don't generalize very well, so they're heavily dependent on how good (how varied as well as how high quality) the training data is

Well it might explain why pro-Claude vs pro-Codex people keep talking past each other on this forum. I see people all the time assuming that anybody who likes Codex must be some sort of bot because of their own biases, but I work almost exclusively in Rust and find Codex extremely competent (and a much better overall engineer), don't trust Claude/Opus at all... but I see in this bench it scores lower on TypeScript etc. than Opus does.

wait why compare 2.6 to 2 instead of to 2.5?

Good question. We missed that release entirely. Our automated model checker only went live 2 months ago so they were manually curated prior to that. I'm adding it now. It'll be live in ~12 hours.

Update: Kimi K2.5 one-shot results are live. It wasn't a noteworthy release compared to K2.6: https://gertlabs.com/?mode=oneshot_coding

I've always been surprised Kimi doesn't get more attention than it does. It's always stood out to me in terms of creativity, quality... has been my favorite model for awhile (but I'm far from an authority)

It’s good, but it’s not quite Claude level. And their API has constant capacity issues.

Price/quality is absolutely bonkers though. I loaded $40 a few weeks/months ago and I haven’t even gone through half of it.

It has long been Claude level since 2.5

Why use China model API from China if there are many independent providers available via Openrouter?

Openrouter will route to china hosted models when there are US hosted providers of the same model. Is there a setting to set your preference or to blacklist providers like alibaba cloud for example?

I use OpenCode and the openrouter provider. From opencode I only select the model like kimi-2.6 and have no way of selecting which cloud hosting will receive my request.

Settings > Guardrails > [your workspace] > Providers + Block provider

to support the companies that open source their models

It's also one of the few models that seem capable of drawing an SVG clock

https://clocks.brianmoore.com/

Interesting that the best performers are all Chinese-made models (DeepSeek and Qwen also perform consistently well). I wonder if there's more focus on vision and illustration in their training, or if something else is leading to their clear lead on this one test.

Is it? In your link it definitely failed to draw the clock.

It redraws it every minute, and some models give quite different results although the prompt is exactly the same.

I'm not really sure how this works, but I stayed on the page for a while, and then it reloaded and all clocks changed. I guess there's either a collection of different clocks generated by models, or maybe they're somehow generated in the real time, but the fact is what you see is not necessarily what I see.

It reruns the prompt every minute.

Dirt cheap on openrouter for how good it is, too. Really hoping that 2.6 carries on that tradition.

Kagi has it as an option in its Assistant thing, where there is naturally a lot of searching and summarizing results. I've liked its output there and in general when asked for prose that isn't in the list/Markdown-heavy "LLM style." It's hard to do a confident comparison, but it's seemed bold in arranging the output to flow well, even when that took surgery on the original doc(s). Sometimes the surgery's needed e.g. to connect related ideas the inputs treated as separate, or to ensure it really replies to the request instead of just dumping info that's somehow related to it.

I remember when the first K2 dropped

It was the best creative writer by some distance

Maybe because it's a bit of like unleashing a chaos monkey on your codebase? I tried it locally (K2.5 72B) and couldn't get anything useful.

Huh, that's not a thing?

The parent poster is probably referring to Kimi-Dev-72B¹, which is a much smaller and older model, while people are probably more familiar with the big and fairly powerful 1100B Kimi-K2.5².

[1] https://huggingface.co/moonshotai/Kimi-Dev-72B

[2] https://huggingface.co/moonshotai/Kimi-K2.5

Yes it was good for its time, but 10 months old now which is a long time ago in this space. It was also a fine-tune (albeit a good one) of Qwen-2.5 72B.

I wish they did more smaller models. Kimi Linear doesn't really count, it was more of a proof of concept thing.

https://huggingface.co/moonshotai/Kimi-K2.6

Is this the same model?

Unsloth quants: https://huggingface.co/unsloth/Kimi-K2.6-GGUF

(work in progress, no gguf files yet, header message saying as much)

A trillion parameters is wild. That's not going to quantize to anything normal folks can run. Even at 1-bit, it's going to be bigger than what a Strix Halo or DGX Spark can run. Though I guess streaming from system RAM and disk makes it feasible to run it locally at <1 token per second, or whatever. GLM 5.1, at 754B parameters, is already beyond any reasonable self-hosting hardware (1-bit quantization is 206GB). Maybe a Mac Studio with 512GB can run them at very low-bit quantizations, also pretty slowly.

A huge dual socket Epyc system used to be able to get to 1TB without difficulty. 16 dimms of 64gb each. Doable for ~$3000. With considerable memory bandwidth.

Our hope these days seems to be that maybe perhaps possibly High Bandwidth Flash works out. Instead of 4, 8, or maybe more for some highest end drives, having many many many dozens of channels of flash.

Ideally that can be very very near to the inference. PCIe 7.0 is 0.5Tb/s at 16x which is obviously nowhere remotely near enough throughout here. The difficulty is sort of that nand has been trying to be super dense, so as you scale channels you would normally tend to scale nand capacity too, and now instead of a 2tb drive you have a 200tb drive prices way beyond consumer means. Still, I think HBF is perhaps the only shot of the most important thing in computing going from mainframe back to consumer, and of course the models are going to balloon again if this dies hit, probably before consumers ever get a chance.

You can't buy 16 64gb dimms for $3000. Go shop memory prices again. But yes an old epyc can run this with no GPU at reasonable speed and if you throw a few GPUs you can get very manageable speed. I run this at home on an old system PCIe4, slow 2400mhz ddr4 ram and still getting about 13tk/sec

Looks like it. This quant ( https://huggingface.co/inferencerlabs/Kimi-K2.6-MLX-3.6bit ) says:

> Q3.6 typically achieves useable accuracy in our coding test and fits within a 512GB memory budget

This one ( https://huggingface.co/mlx-community/Kimi-K2.6-MoE-Smart-Qua... ) though says it fits on a 192GB mac:

> M3/M4 Ultra 192GB+ (fits in ~150GB)

Quite curious how well real usage will back the benchmarks, because even if it's only Opus ballpark, open weights Opus ballpark is seismic.

Huh, so the metadata says 1.1 trillion parameters, each 32 or 16 bits.

But the files are only roughly 640GB in size (~10GB * 64 files, slightly less in fact). Shouldn't they be closer to 2.2TB?

The bulk of Kimi-K2.6's parameters are stored with 4 bits per weight, not 16 or 32. There are a few parameters that are stored with higher precision, but they make up only a fraction of the total parameters.

Huh, cool. I guess that makes a lot of sense with all the success the quantization people have been having.

So am I misunderstanding "Tensor type F32 · I32 · BF16" or is it just tagged wrong?

The description specifically says:

"Kimi-K2.6 adopts the same native int4 quantization method as Kimi-K2-Thinking."

Has anyone here used Kimi for actual work?

I tried it once, although it looks amazing on benchmarks, my experience was just okay-ish.

On the other hand, Qwen 3.6 is really good. It’s still not close to Opus, but it’s easily on par with Sonnet.

Yes. You’re using Kimi if you use the composer-2 model in cursor. It’s great. Plan in state of the art. Execute in composer-2

Before GLM-5.1, I was going back and forth between Opus 4.5 and Kimi 4.5 and having very good results with Kimi.

I've used Kimi K2.5 when I run out of Codex quota. It does small and medium things OK. But if I work on complex things, I'll later have to spend two days cleaning up the mess with Codex. Hopefully 2.6 does better.

Wow, if the benchmarks checkout with the vibes, this could almost be like a Deepseek moment with Chinese AI now being neck and neck with SOTA US lab made models

With the previous generation? Yes. With 10T mythos-level models? Not even close.

The psyop continues. Mythos until it’s released is vaporware. Notice how you can try kimi 2.6. Where is the same for mythos?

It's been released to "select partners".

At this point it seems more like the result of a psyop to presume that a new anthropic model should be considered vaporware until released.

I've got a 12T model on my machine, built it myself. It's called Mytho. Too dangerous to even release a fact sheet about it. It can hack into the mainframe, enhance ultra-compressed images, grow your hair back, and make people fall in love with you.

Mythos isn't the current generation, it's literally vaporware.

I doubt it's literal vaporware. It's likely just a variant of whatever model they just generally released with some fancy prompt and a highr quant.

According to the benchmarks, you are wrong. It is on track and slightly above some sota. Just the benchmarks speaking there, they can be/are gamed by all big model labs including domestic.

There's no public data about Mytho.

That's because it would be too dangerous to release.

10T? Impossible! They told us the training run was under 10^26 flops.

Mythos doesnt exist

mythos is a mythos

mythos is vaporware right now, what are you talking about?

Well, the model that a lot of people have been given access to and are reporting about on twitter?

I have been testing it in my app all morning, and the results line up with 4.6 Sonnet. This is just a "vibe" feeling with no real testing. I'm glad we have some real competition to the "frontier" models.

it feels like between K2.6 and GLM5.1 we have Sonnet level intelligence at roughly Haiku level pricing. Which is great.

I'm hoping that Anthropic will be able to release an updated Haiku soon and they really need something that is 1/3-1/5 the price of Haiku to compete with the truly cheaper models (Gemma-4 is really good at this range).

In my tests[0] it does only slightly better than Kimi K2.5.

Kimi K2.6 seems to struggle most with puzzle/domain-specific and trick-style exactness tasks, where it shows frequent instruction misses and wrong-answer failures.

It is probably a great coding model, but a bit less intelligent overall than SOTAs

[0]: https://aibenchy.com/compare/moonshotai-kimi-k2-6-medium/moo...

I tried it on openrouter and set max tokens to 8192, and every response is truncated, even in non-thinking mode. Maybe there's an issue with the deployment, but in your link also shows it generates tons of output tokens.

Oh yeah, I just noticed, like 3x the reasoning tokens.

I pray the benchmark figures are true so I can stop paying Anthropic after screwing me over this quarter by dumbing down their models, making usage quotas ridiculously small, and demanding KYC paperwork.

Absolutely. Thing is, I'd actually rather take a worse model than Anthropic, so long as it's consistent. Like, a model that can successfully do well for 80% of tasks is much better than Anthropic that some days will be 90% other 60%.

When you have a consistent model, you can incorporate fixes/prompts into your workflow to make it behave better. But this, always having to guess if Anthropic has quantised the model today, wastes so much time and effort.

Codex has a lot better limits, and 5.5 will be out soon

Anthropic has done horrible PR and investors should be livid.

My theory is they pushed retail off their systems to make room for their new corporate fat cat clients. In which case, they'll do just fine.

> dumbing down their models,

This should be so easy to prove if it were true. Yet there is none of it, just vibes.

Still, your other two points are completely valid. The opaqueness of usage quotas is a scam, within a single month for a single model it can differ by more than 2x. And this indeed has been proven.

> This should be so easy to prove if it were true.

https://github.com/anthropics/claude-code/issues/42796

https://scortier.substack.com/p/claude-code-drama-6852-sessi...

First link is about the harness, Claude Code, defaulting to less thinking over time. This isn't "the model getting worse".

Second link is just a discussion of the first link.

I often wonder if in the future, the same way early computers used to take up an entire room but now fit in your pocket, if in the future the equivalent of a data center will be a single physical device like a phone nowadays. And if that’s the case, would it happen much quicker since technology has been speeding up year by year?

> And if that’s the case, would it happen much quicker since technology has been speeding up year by year?

I wouldn't expect this.

Historically we've had a roughly exponential rate of shrinkage. If we keep that same exponential going, we should expect the amount of time to shrink "room full of compute" to "pocket full of compute" to be equal.

And recently we've fallen behind that exponential rate of shrinkage. And this is rather expected because exponentials are basically never sustainable rates of growth.

I still expect that technological progress is getting faster year by year, and that we're still shrinking compute, but that's not necessarily enough for the next shrinking to take less time than when we had exponential progress on shrinking.

There’s some early work being done here by companies looking at making LLM ASICS like Taalas (HC1 gets 17k t/s for llama 8b - currently at 2.5kW which is closer to a single server, but this is their first chip).

There’s other options like photonic computing which might be able to reduce power significantly but are still in research as far as I can tell. Because so much money is invested in AI & traditional gpu inference is so power hungry, I would expect significant improvements in this space quickly.

K2.5 was already pretty decent so I would try this. Starting at $15/month: https://www.kimi.com/membership/pricing

edit: Note that you can run it yourself with sufficient resources (e.g., companies), or access it from other providers too: https://openrouter.ai/moonshotai/kimi-k2.6/providers

What's the privacy/data security like? I can't find that on that page.

Edit: found it.

> We may use your Content to operate, maintain, improve, and develop the Services, to comply with legal obligations, to enforce our policies, and to ensure security. You may opt out of allowing your Content to be used for model improvement and research purposes by contacting us at membership@moonshot.ai. We will honor your choice in accordance with applicable law.

Section 3 of https://www.kimi.com/user/agreement/modelUse?version=v2

> We will honor your choice in accordance with applicable law.

So in other words only if you can point to a local law which requires them to comply with the opt out?

most laws enforce agreements.

Yes... but the agreement only says they won't train on your data if the law is already preventing them from doing so.

You really rely on ToS from Anthropic/OpenAI to know if they use your prompts or not? It's on their servers, why wouldn't they use our data?

Antropic and OpenAI are used by US businesses and government and they are audited and under contracts.

If it's discovered they trained on data they shouldn't have had it will be the end of their business.

On the other hand, good luck suing a Chinese company.

Not at all, Google/Meta... got caught all the time, where do you see it's the end of their business?

Yup, they train on your inputs and OpenRouter is complicit by claiming that Moonshot's ToS says that they don't. Contacted OpenRouter about this a while ago and was met with silence because it's bad for their business to stop lying about it.

"sufficient resources" is going to be a lot of resources. I doubt this will run on even something like a Strix Halo or DGX Spark, even at 1-bit quantization. You'll need a 256GB or 512GB Mac Studio, or a monster GPU situation, to run it locally, I think, though quantized versions aren't showing up yet, to be sure.

How are the usage limits compared to Anthropic?

Anthropic has the worst usage limits in the industry

gemini is worse imo

You're correct, Gemini chat limits are a joke at their chapest paid tier compared to both Claude and GPT. Especially crazy when you consider Gemini 3 Pro is more than twice as cheap as Opus 4.6 on the API. It's hard to run into pure chat limits on Claude even if you only use Opus on the cheapest tier, whereas with Gemini it's easy to hit.

Not sure about coding usage, Google being weird about these things I could see that quota being separate.

Are there any coding plans for this? (aka no token limit, just api call limit). Recently my account failed to be billed for GLM on z.ai and my subscription expired because of this... the pricing for GLM went through the roof in recent months, though...

Kimi has their own subscription that works basically the same as all the others.

https://www.kimi.com/code

At $19/month, hard to see why I want to use Kimi over Claude.

Claude usage at $20 is basically unusable for serious work. I haven't used Kimi but I'd have to imagine they're offering a good deal more usage for the same price.

Because Opus on $20 CC is a joke. The $19 plan on Kimi has actually workable usage limits.

for similar plan i think claude costs like $100 a month?

They tick all the boxes I care about – desktop & mobile app, cli – but for the same price I might as well just go for the leading providers.

You can use $20 pro plan on Ollama or $10 one on OpenCode Go. Both has Kimi 2.6 live. https://opencode.ai/go https://ollama.com/pricing

Really excited to try this one, I've been using kimi 2.5 for design and it's really good but borderline useless on backend/advanced tasks.

Also discovered that using OpenCode instead of the kimi cli, really hurts the model performance (2.5).

I have a subscription through work, I've been trialing it, so far it looks on par, if not better, than opus.

I tried it out with my normal mixed-up wolf, goat, cabbage problem and it couldn't solve it. Sonnet 4.6 also can't, but Opus 4.7 has no problems.

Details here [0]

[0] https://techstackups.com/comparisons/kimi-2.6-vs-opus-4.7-an...

wow - $0.95 input/$4 output. If its anywhere near opus 4.6 that's incredible.

This should erase any doubt that AI Labs are making $$$ on API inference.

Kimi 2.5 (which this is based on) is served at $0.44 input / $2 output by a ton of different providers on OpenRouter, 2.6 will certainly be similar.

That's about 11X less than Opus for similar smarts.

It’s worth noting that the US is very behind on energy infra and that might affect the cost calculations since data centers are electricity guzzlers. Also, not sure if CN has completely switched off Nvidia or still using them for training.

Famously, OpenAI and Anthropic are devoted to increasing efficiency before scaling up resource usage.

How does it erase any doubt? You’re implying Chinese things can’t be actually cheaper to produce than American which is laughable

Most of those inference providers are American, and China is actually at a disadvantage here because of export restrictions - US companies are using newer and more efficient chips.

If it’s newer and efficient then why is the api more expensive?

I'm pretty Kimi is what Cursor uses for their "composer 2" model. Works pretty good as a fallback when Claude runs out, but definitely a downgrade.

It's a Kimi K2.5 finetune, there was some drama about this a few weeks ago.

What was the drama about?

They were not open about the fact that it was trained on Kimi K2.5. This link explains it better than I can:

Cursor seemingly went out of their way to not mention that they were actually running Kimi K2.5 and essentially by that omission made it seem like they had made their own model. They added a note to a blog post about using it at some point and then when they wrote a new one they conveniently left it out again.

That's at least what I perceived as "the drama".

Damn it, they stopped offering Kimmmmy. Their sales ai agent which allowed you to bargain for lower subscription prices.

Beats Opus and Open Source?

I really hope this holds true in real world use cases as well and not only benchmarks. Congrats to Kimi team!

K2.6-code-preview was a minor, but noticeable jump, especially in a long running testing task and prior Moonshot releases have been the only models that I'd consider a suitably competitive replacement for Anthropic models. The way they approach tool calls, task inference and adherence is far closer than any other providers output, similar to how GLM models map far more closely to OpenAIs releases. Whether task adherence, task assessment, task evaluation or task inference, K2.5 got closer to Opus 4.5 than any other model (but was still behind overall).

I will have to test this full release of K2.6 but could see it serve as a very good overall drop-in replacement for Opus 4.5 and Opus 4.6 at 200k across the vast majority of tasks.

I will say however that Opus 4.7 Max 1M has been a very significant jump in performance for me, especially in tasks beyond 120k token where I'd argue it is now the most reliable model in continued task adherence and tool calling without compaction. Ironically, my initial experience was less than pleasant as on XHigh I found task adherence to have regressed even with less than 1/10th of the context window having been used.

Am very interested in K2.6s compaction strategy (which appears to be very simply all things considered) and how it performs beyond 100k tokens. As it stands, only OpenAI models have made compaction for long running tasks work well, though overall, GPT-5.4 is still inferior in my tests regardless of context window over other models such as Opus 4.6 1m and Opus 4.7 1m. Haven't gotten around to testing Opus 4.7 200k and will have to do this to properly assess K2.6 fairly, but I'd be very surprised if K2.6 truly beat Opus 4.7 200k given the jump I have experienced.

Beats opus 4.6! They missed claiming the frontier by a few days.

While I'm skeptical of any "beats opus" claims (many were said, none turned out to be true), I still think it's insane that we can now run close-to-SotA models locally on ~100k worth of hardware, for a small team, and be 100% sure that the data stays local. Should be a no-brainer for teams that work in areas where privacy matters.

Even the smaller quantized models which can run on consumer hardware pack in an almost unfathomable amount of knowledge. I don't think I expected to be able to run a 'local Google' in my lifetime before the LLM boom.

I'm extremely curious how these models learn to pack a lossily-compressed representation of the entire Internet (more or less) into a few hundred billion parameters. like, what's the ontology?

I think this one is only about 600GB VRAM usage, so it could fit on two mac studios with 512GB vram each. That would have costed (albeit no longer available) something like less than 20k.

Yeah, but that's personal use at best, not much agentic anything happening on that hardware. Macs are great for small models at small-medium context lengths, but at > 64k (something very common with agentic usage) it struggles and slows down a lot.

The ~100k hardware is suitable for multi-user, small team usage. That's what you'd use for actual work in reasonable timeframes. For personal use, sure macs could work.

True, but I think for local models, we are mostly considering personal usage.

You could run it with SSD offload, earlier experiments with Kimi 2.5 on M5 hardware had it running at 2 tok/s. K2.6 has a similar amount of total and active parameters.

Yeah... I would definitely call 2t/s unusable. For simple chats, I'd want at least 15 t/s. For agentic coding (which this model is advertised for), I'd want good prefill performance as well.

That's just throwing money away. The performance with large context would have been unusable especially if you need to serve more then a single person.

Opus is clearly a sidegrade meant to help Anthropic manage cost, so I would say they may have it if it actually beats 4.6

Could be right. I just noticed my feed is absent the usual flood of posts demoing the new hotness on 3D modeling, game design and SVG drawings of animals on vehicles.

It doesn't beat Opus 4.6, no way, don't be fooled by benchmarks.

I really wish some of these very-long-horizon runs were themselves open sourced (open released open access). Have the harness setup to do git committing automatically of the transcript and code, offload the git commit message making. Release it all.

This sounds so so so cool. It would be so amazing to see this unfurl:

> Kimi K2.6 successfully downloaded and deployed the Qwen3.5-0.8B model locally on a Mac. By implementing and optimizing model inference in Zig—a highly niche programming language—it demonstrated exceptional out-of-distribution generalization. Across 4,000+ tool calls, over 12 hours of continuous execution, and 14 iterations, Kimi K2.6 dramatically improved throughput from ~15 to ~193 tokens/sec, ultimately achieving speeds ~20% faster than LM Studio.

Running it through opencode to their API and... it definitely seems like it's "overthinking" -- watching the thought process, it's been going for pages and pages and pages diagnosing and "thinking" things through... without doing anything. Sitting at 50k+ output tokens used now just going in thought circles, complete analysis paralysis.

Might be a configuration or prompt issue. I guess I'll wait and see, but I can't get use out of this now.

Had the same experience using it for a refactor of a 3k LOC monolith via the Pi harness and OpenRouter. After burning through $8 worth of tokens it left the code in a broken state, the "thoughts" were full of loops where it would edit the monolith, then refer back to the original file, not finding it and then overwriting its changes with "git checkout --"

It's probably bad harness. I had a similar bad experience with qwen max yesterday also through opencode.

In the past I tried Kimi thru Claude code I might try that again

I think this kind of overthinking is an extremely common pattern in the Chinese models. GLM's models are also very much like this.

The choice of example task for Long-Horizon Coding is a bit spooky if you squint, since it's nearing the territory of LLMs improving themselves.

Added support for Kimi in https://github.com/raine/claude-code-proxy and it does appear to work surprisingly well with Claude Code, although the usage limit for the entry tier doesn't seem as generous as I'd have expected.

If the benchmarks are private, how do we reproduce the results? I looked up the Humanity's Last Exam (https://agi.safe.ai/) this model uses and I can't seem to access it.

You can request access here: https://huggingface.co/datasets/cais/hle

The test data is purposely difficult to access to reduce the chance of leaking it into the training dataset.

Am I being paranoid in questioning whether the CPC would have something to gain by monitoring coding sessions with Chinese coding AI models? Coding models receive snippets of our intellectual property all day long. It's a bit of a gold mine, no?

I think you should worry more about NSA, FBI, ICE and other 3 letter US agencies monitoring your sessions

There's nothing anyone can do about state-level espionage anywhere, using any cloud-hosted service. That being said, there is a very big difference between the legal situation in the United States vs. China. Chinese internet companies are required to have CPC interaction and since the rule of law does not strictly exist in China, the state can compel surveillance cooperation regardless of what might be written down. If a three-letter agency is compelling Anthropic to open up its queries for inspection, that kind of surveillance would be authorized by law and if Anthropic violated the law in cooperating, they would suffer the consequences in civil court. Maybe not immediately, but at least the possibility exists.

In China, there's no recourse at all. Surveillance must be presumed.

> the rule of law does not strictly exist in China, the state can compel surveillance cooperation regardless of what might be written down

While I agree that China is obviously worse in this regard, it's naive to claim this is unique to China, when literally a couple of months ago the US got into a fight with Anthropic about them not removing safeguards which were already just enforcing the letter of the law.

Rule of law in the US - are you kidding yourself?

When American citizens are being gunned down in public on cameras by US federal government agents, you are telling me that the US follows the rule of law?

Before you start to offer more propaganda, just tell me where is the killer of Renée Good, has that killer been arrested or charged yet? Keep your censored version of rule of law to yourself and your kids.

oh, btw, the current US President did got convicted for criminal offences, he walked away for free just because he got elected as the president. nice rule of law! what did he do recently - authorised illegal war against another country in which over 100+ school children got killed. Surely your fancy US rule of law is going to do something about this?

Are there any protections from industrial espionage when using Anthropic, Cursor, Gemini, or OpenAI?

There are legal protections, and those companies have more to lose by breaking those laws than following them. Same probably not true for Chinese companies.

Legal protection, only if you're a billionaire and US citizen, for everyone else there is no protection.

Does US actually follow laws? They literally kidnapped head of another state and bombed another state and you are expecting legal protection from them?

You don't have to be a US citizen or live in the US to file a lawsuit against an American company in the US court system. Federal courts explicitly allow it under the "alienate jurisdiction" clause.

Exciting benchmarks if true. What kind of hardware do they typically run these benchmarks on? Apologies if my terminology is off, but I assume they're using an unquantized version that wouldn't run on even the beefiest MacBook?

With agents running at the scale and for an extended period. Surely they would need to pay for external services like APIs, compute, data. Would everything be based off subscriptions or API usage?

> Agent Swarms, Elevated: Match 100 Jobs and Generate 100 Tailored Resumes

Model seems quite capable, but this use-case is just yikes. As if interviewing isn't already a hellscape.

This kimi website, it looks like a stylesheet from the 90's. They could learn a thing or two about typeface design. Steve Jobs would be incensed at this.

I prefer a website that has the first page of text visible almost immediately, with no glitches when fonts load, tbh.

Here I analyze the same linenoise PR with Kimi K2.6, Opus, GPT. https://www.youtube.com/watch?v=pJ11diFOjqo

Unfortunately the generation of the English audio track is work in progress and takes a few hours, but the subtitles can already be translated from Italian to English.

TLDR: It works well for the use case I tested it against. Will do more testing in the future.

The modified MIT clause is sneakier than people think. Hit 100M users or $20M a month and you have to slap "Kimi K2.6" on your UI. That covers any consumer app worth building. Not really open, more like free until you matter. Llama pulled the same move

The threshold for "worth building" is much lower than that for a lot of people.

Attribution is a fair clause in opensource. What is the problem? You are making 20M$ a month thanks to their free work.

Worth building with VC capital maybe. A small team putting together an app that pulled in $20M per year should be pretty pleased with that.

I'll definitely put this into the "good problem to have" category.

if you reach that numbers, kimi would be your least of worries

in what way does this restrict how you are able to use the model?

And the Kimi team broke the Anthropic ToS by training off Opus outputs and… nothing happened?

Nobody cares, nor should they. Anthropic broke nearly every ToS of every website that they scraped data from. The AI robber barons just want to monopolize intellectual property violations, and I'm gonna cheer on any robin hoods that take it back from them.

If only their API wasn't tied to a Google or phone login...

If it's open then there will be multiple providers. I see it is on OpenRouter now.

I'm going to experiment with this, but unless it's insanely more efficient in token usage than anything else I've tried, the only way to keep costs more or less acceptable is through a subscription.

Why use "their API"? It is an open model, use any provider on OpenRouter

Because sometimes (a lot of the time in my experience) third-party providers and inference engines fail to implement the model correctly in ways that are sometimes very subtle and not obvious.

Deepinfra for example is not preserving thinking correctly for GLM5.1, even though they are for GLM5. This is one of the more obvious issues that crop up.

Yeah some of the providers on Openrouter correctly list what quantization they offer. And some refuse to say. OR should kick them off their platform if they want to be secretive.

Does it run on Nvidia or Huawei?

isnt this better than qwen?

We'll have to wait for the results on Artificial analysis

(commented on the wrong thread, HN doesn't let me delete it :( )

They're comparing to Opus 4.6, not 4.5. It was Anthropic's best public model up until last week.

Some people would say it's still Anthropic's best public model!

Yeah, I noticed that, HN doesn't let me delete my comment.

The other release, Qwen-3.6-Max is the one comparing it to 4.5

Reply on news.ycombinator.com