kamranjon

“QuickJS is to this PSP what V8 is to Node: the host hands it a strike API surface and a ui API surface, and the same openstrike.js bundle boots against them on every target.”

This comparison seems backwards? Also I have no idea what it means - it seems like really strange phrasing and I can’t seem to make sense of it. Does anyone understand what was trying to be conveyed here?

Hearing a world leader talk about open source is huge. I think it highlights pretty starkly the difference in approach between the U.S. and China and will likely in time be seen as a pretty historic statement.

Where did you hear about the deepseek release? Would love to follow the same source.

“Alongside Inkling we are sharing a preview of Inkling-Small, a 276B-parameter Mixture-of-Experts model (12B active, vs. 41B for Inkling) with a different performance/latency trade-off.”

Buried at the end there is the details I was most interested in - a possible competitor for DeepSeek V4 Flash? Excitedly awaiting the release of the weights for this one.

70% seems a little extreme? Here is a helpful mobile coverage map: https://www.fcc.gov/BroadbandData/MobileMaps/mobile-map

A few years ago I actually traveled all around the US and worked remotely with a 4g unlimited data sim card and was pretty amazed at the coverage, only very few places was I left without coverage. I'm guessing it's even better now?

I am traveling in Europe currently and got a Saily SIM card - 5g coverage is really good and seems to be expanding fast - a small village I visited 2 years ago that had no cell coverage at all I was getting 250mbps download, faster than than the wired internet at most of the places I was visiting. This prompted me to actually look into Starlink speeds and apparently 5g is generally faster than Starlink...

This made me wonder if Starlink is a much more niche product than I thought - I actually thought it was significantly faster than cell data - but it seems it's just for folks in super remote areas at this point?

Doesn’t this suggest you aren’t properly running the model?

Curious why you find that’s a useful test since it seems to be solely measuring training data memorization, something you’d expect to degrade from quantization.

When new models are released (I realized this is qwen 3.6 but the quant is novel) - it takes a few days for the kinks to get worked out - you’ll likely have better luck if you give it a few days and try again.

Yea I don’t use sub-agent style workflows. I often use planning patterns and generate markdown for really complex tasks, but never got into the sub-agent thing. I am likely closer to AI-assisted, though I am heavily using pi coding agent and my editor is basically just for viewing files and changes now.

DeepSeek V4 Flash with DwarfStar: https://github.com/antirez/ds4

The 2 bit quants are really good. I have a lot of memory so I can squeeze it all in at ~80gb.

I’m using DeepSeek V4 Flash on 128gb mbp - it’s a bit different using a 200b+ param model. It’s MoE so performance is acceptable. It will still malform a tool call every now and then, but the capabilities are so far ahead anything else that the majority of the time it works really well and solves really complex problems.

After using a highly capable 2-bit quant as my daily driver for months now, I get pretty excited about releases like this. After a few days for the kinks to be worked out, I’ll be excited to try it.

I haven't read No Country yet - I think at the time it was Suttree which is definitely in my top 5 - I also read The Road recently and was pretty blown away, really quick read and very meticulously structured, I loved it. I'll make a point to read No Country, I heard it was originally written as a screenplay and so is less descriptive than his other work.

I actually read the first book and it was so poorly written it made me wonder if I should continue, because I did find the general story quite engaging. I've heard it gets better/tighter in subsequent books, but it was the only King i've ever read so wasn't sure if he just had a sort of sloppy style (he did write pretty prolifically). I also read it following a Cormac McCarthy book so that might have lead me to believe it was sloppier than it deserved.

Oh that’s great - I’m generally more interested in single machine use cases but the RDMA stuff is super interesting.

I do wonder if you are planning to eventually delegate some of the merging responsibilities? I see some other projects follow a similar pattern when they grow to a certain size.

Where do these estimates come from?

I actually have asked many people, because it's something that's frequently on my mind (I had a bit of a revival with reading and do have some strong opinions about how it's helped me). Are you suggesting that you listen to audio books while sitting silently in a chair at your home? That's great! I just don't think that is true for the majority of people. The benefit by and large, as far as I can gather, for most people, is being able to consume something while doing something else. And like I said, that's fine - and I use it myself when exercising - but I just don't consider it reading - and think of it as a separate type of activity that doesn't require the level of engagement that reading a physical book does.

Some of the most exciting engineering work is happening in the DS4 repo - and I'm watching it almost like a sports game.

When the DSpark paper came out[1] the next day we had folks attempting to implement, working together, validating their failures. Eventually their work being synthesized into a PR[1] that admits performance is not ideal. Something antirez alluded to in one of his videos (speculative decoding is a great boon, but mostly for large labs hosting and serving many requests at once, and maybe not so effective for local inference).

There's recent work into "directional steering"[3] that has made it's way into per-session directional steering overrides thanks to audreyt[4].

There's support for the new Hy3[5] model also thanks to audreyt[6].

There's Pre-M5 optimizations[7] in the queue thanks to ivanfioravanti who also helped with some of the initial M5 optimizations.

I haven't watched a repo like this since llama.cpp and whisper.cpp in the early days (though llama.cpp is pretty exciting right now with the SYCL improvements that are flowing in for the new Intel GPUs).

The DS4 repo is a really interesting place to watch folks who heavily code with agents collaborate together in a way that seems pretty effective. I've been really enjoying it.

[1]: https://arxiv.org/abs/2607.05147

[2]: https://github.com/antirez/ds4/pull/502

[3]: https://arxiv.org/html/2406.00045v2

[4]: https://github.com/antirez/ds4/pull/148

[5]: https://hy.tencent.com/research/hy3

[6]: https://github.com/antirez/ds4/pull/523

[7]: https://github.com/antirez/ds4/pull/555

In your examples you are always doing something else while listening to an audio book. This was the authors point. Reading is the activity - and I definitely agree - I still listen to audio books, but for a different purpose, to avoid boredom when exercising, etc but I don’t consider this reading. I really think most people can learn quite a bit, especially in this age of instant gratification, by sitting down and performing just the task of reading. For me it is a huge boon for my mental health to engage in this way. I think you could do it with audio books too, but I don’t know a single person who engages with them in that solitary way, it is always an addition to some other activity.

I’ve actually been really interested in Minimax M3 - seems like it flew under the radar but size wise might actually be runnable for local inference with a footprint somewhere between Deepseek V4 flash and pro.

Has anyone used the new Minimax M3 model? I’m curious how it compares with Deepseek V4 and GLM 5.2 and other larger open weights models.

The Minimax paper was published in June 2026 coinciding with the Minimax M3 release - I’m not sure how the repo you posted here could have been an implementation of Minimax sparse attention when it was updated over a year ago?

I dunno I use LMStudio pretty regularly and the MLX folks and the community usually have MLX versions of new model releases up within a day or two.

Antirez has a GLM 5.2 branch now in dwarfstar: https://github.com/antirez/ds4/tree/glm5.2

It heavily utilizes ssd streaming from my understanding and I think he mentioned getting some semi usable speeds on a 128gb m5 mbp.

Not sure it’s a blog post really but I saw a video of Rutger Bregman talking about moral ambition and it directly led to me applying at a non-profit that was better aligned with my own ideals and leaving my job at a pretty large and morally reprehensible corporation.

Wanted to share Antfly which I think serves a similar niche:

https://antfly.io/

https://github.com/antflydb/antfly

They’ve put a lot of effort into optimizing the local llm pipelines and I have a lot of faith in the devs working on it.

It’d be real funny if this was just GLM 5.2 trained on Cursor data

This would be a pretty cool addition to the duckdb HNSW search project I found on here some time ago: https://github.com/jasonjmcghee/portable-hnsw

What I think is really cool is that the search happens using http range queries across statically hosted parquet files.

I think things like this could bloom into a relatively open and distributed search ecosystem that isn’t controlled by major corporations.

One of the big benefits I’ve heard about for the framework, especially with the Noctua cooler, is that it’s virtually silent.

In case it saves anyone some time (from the article): "The AMD Ryzen AI Max+ 395(Strix Halo) processor has been available since Spring 2025 and the Halo doesn’t offer anything new on that front."

It has the same 256 GB/s memory bandwidth limit as every board previously, not sure why this is even being released right now as if it's some new fangled thing - you can go get a Framework Desktop for roughly the same price or a GMKtec EVO-X2 for a bit cheaper.