mekpro

I did use deepseek via API and it can easily cost 3$ per day for my workload. On the other hand Luna 6 on 20$ subscription from OpenAI cover by usage for whole month. However, `ccusage codex` command suggested that my Luna6 token cost is around ~100$ monthly.

``` or by 5 opencode subscriptions, and have $800 of ds4.1-flash for $200. I'm skeptical that the average dev will manage to burn through the equivalent of $800 of ds4.1-flash if they spend even a bit of time optimizing their token usage. ```

This is my point. For most developer, subscription model (with subsidize) is preferable. And what I'm wondering is, as suggestted by very cheap cost-per-token from DeepInfer, why don't Deepseek persued subscription market as well ? What make subscription not make sense for them to provide ?

Even though their API price is cheap compared to OpenAI/Anthropic, I still can’t “afford” to use their API because the subscriptions from OAI/Anthropic provide far higher value and peace of mind. I know for sure how much I need to pay each month. I believe most developers fall into this category as well, where $20/$100/$200 from Anthropic/OAI provides more value.

This raises the question: why doesn’t DeepSeek provide a subscription service as well? Is it a better move to have a single pricing model like they currently do?

Why they are not even benchmark model against Anthropic or anybody ?

It’s interesting in terms of the economics of development.

This product takes advantage of advancements in lithium-ion battery technology driven by phones and cars, benefiting from improvements in both performance and cost at scale without requiring extensive additional research. Looking forward to trying it.

why so much hate on this ? Meta releases open Model and it is competitive and confirms the direction, what could we ask more !!

codex spark is not large model though, much weaker than standard model.

We need more coding benchmark score. Not sure that winning terminalbench 2.1 alone is a clear win over Fable/Mythos yet.

source ?

API server is not hard problem and not make sense for indefinite postpone. I think the more likely explanation is model quality.

Too bad for Meta, and very sad day Llama.

The technical report is very detailed and would 'reinforcement learning' of future researchers, Thanks Microsoft!

Yes, 300 MW from SpaceX helps a lot, but I think that’s mainly to support Opus demand, which has grown faster than expected. If Mythos is roughly 5× more expensive to serve than Opus, as the pricing suggests, then 300 MW is nowhere near enough to enable large-scale deployment of Mythos.

As an ordinary developer who relies on a $20–$200/month subscription, I feel disappointed by the release of a paper describing a model that I can’t actually use.

It’s clear that Anthropic has run out of the compute capacity needed to serve Mythos publicly.

They’re using security concerns to mask their inability to deliver the model at scale, while still trying to maintain their lead over OpenAI. As a result, they’ve chosen to release it privately under the banner of an “ethical” rollout.

Opus is definitely in its own league. I use Kimi/Gemini-cli code regularly to save cost and from my experience, Kimi 2.5 is more solid than Gemini Flash 3.0 for coding. While Gemini Flash 3.0 is generally faster, it usually break the syntax and skip important prompt. Kimi 2.5 can write very good code and can plan very well.

Except that, In OpenRouter, Deepseek always maintain in Top 10 Ranking. Although I did not use it personally, i believe that their main advantage over other model is price/performance.

I think the opposite. Having NVIDIA investing in TSMC's bleeding-edge process node should benefit Apple rather than disadvantage.

It means that Apple doesn't have to be sole investor in latest node development which is more harder to justify, especially in the year where smartphone upgrade cycle is slowdown. Having NVIDIA (and AI boom) in the picture should help Apple reduce CAPEX for their semi-conductor investment.

They are so beautiful that i dont want any of these been stole by AI.

How this improvement translate into real world agentic coding task ?

i got 70 token/s on m4 max

try enable flash attention and offload all layer to GPU

Is this limit will also count together with Claude Chat ?

you can easily reach 50$ per day. by force switching model to opus /model opus it will continue to use opus eventhough there is a warning about approaching limit.

i found opus is significantly more capable in coding than sonnet, especcially for the task that is poorly defined, thinking mode can fulfill alot of missing detail and you just need to edit a little before let it code.

Just refactored 1000 lines of Claude Code generated to 500 lines with Gemini Pro 2.5 ! Very impressed by the overall agentic experience and model performance.

To professionals in the field, I have a question: what jobs, positions, and companies are in need of CUDA engineers? My current understanding is that while many companies use CUDA's by-products (like PyTorch), direct CUDA development seems less prevalent. I'm therefore seeking to identify more companies and roles that heavily rely on CUDA.

it can use tool to explore directory like ls grep out of the box.

Gemma 3 is way way better than Llama 4. I think Meta will start to lose its position in LLM mindshare. Another weakness of Llama 4 is its model size that is too large (even though it can run fast with MoE), which greatly limits the applicable users to a small percentage of enthusiasts who have enough GPU VRAM. Meanwhile, Gemma 3 is widely usable across all hardware sizes.

Google is also the only company that has had their own AI hardware that's worked (TPU). This could lead to more cost-effective training + inference and hence better AI.

Also, they open-model gemma-3 is very competitive for its size and actually beats llama-3 from Meta. Not to mention that OpenAI doesn't offer anything open anymore.

Still a big bubble considered that the price go up 244% compared to 6 months ago.

As a quick estimation, the size of q4 quantized model usually be around 60-70% of the model's parameter. You can preciselly check the quantized model size from .gguf files hosted in huggingface.

Zero to One