DeepSeek V3.2 which uses DSA only (sparse attention, but without compression from HCA and CSA) is a smaller model but uses 10x more memory at 1M context window compared to DS V4 Pro.

Also, I have to say, DeepSeek's API has a very good cache hit rate. With the same workload, I see ~80% KV cache hit rate with the DS API vs ~50% with the major western inference providers for open weight models.

Inference stack efficiency: Many of these providers take off the shelf sglang / vllm / trtllm and hope for the best. Meanwhile DeepSeek team is known for pushing the boundary of optimizations.

Now, sglang and vllm are great pieces of software, but take DeepSeek's Sparse Attention (DSA). Introduced 1.5 years ago (https://arxiv.org/abs/2512.02556), used by DeepSeek 3.2, GLM 5, DeepSeek V4. Only now is it slowly strating to get optimized in the major inference engines: (https://github.com/sgl-project/sglang/issues/19380 https://github.com/sgl-project/sglang/pull/22851 etc.). Of course, DS V4 adds extra optimizations into the model architecture on top of DSA, and those will take more time to be taken full advantage of by the open source inference engines.

Privacy: Betting that people will pay extra for inference hosted outside China. This is especially true with DeepSeek, because DeepSeek is transparent about using API data for model improvements.

And few other things (scale (matters a lot for MoEs), reliability, soft enterprise lock in, etc.)

---

There is also, likely, tacit collusion at play here. Look at GLM 5 and GLM 5.1 prices. GLM 5 and 5.1 cost the same to run, but providers decided to charge much more for 5.1 because it is much better model, and because Z.AI raised their price as well.

Comments URL: https://news.ycombinator.com/item?id=48033664

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=47931492

Points: 1

# Comments: 0

Model makers (both open and closed weight) typically publish benchmarks against other models and when they do not, people rightfully call them out.

Including comparison against "other OSS engine" is just not helpful (what if it's a sandbagged baseline like HF Transformers?)

Bechmarks from InferenceX (they do not have apples-to-apples setups to compare the different engines for whatever reason): https://inferencex.semianalysis.com/inference?i_hc=1&g_model...

I find it odd that sglang, vLLM, TRTLLM don't seem to want to publish benchmarks comparing each other. They used to, but now there seems to be some unspoken rule against it.

At least we get comparison against "other OSS engine" this time, but that could be HF's Transformers as well :)

Comments URL: https://news.ycombinator.com/item?id=47902025

Points: 2

# Comments: 0

This model is 8 times cheaper than Gemini for 1K images. Gemini is extremely overpriced.

1K image with Gemini is roughly $0.08 and only $0.01 with GPT Image.

Comments URL: https://news.ycombinator.com/item?id=47832222

Points: 485

# Comments: 348

Comments URL: https://news.ycombinator.com/item?id=47798042

Points: 25

# Comments: 11

Comments URL: https://news.ycombinator.com/item?id=47796950

Points: 6

# Comments: 2

Comments URL: https://news.ycombinator.com/item?id=47672120

Points: 3

# Comments: 0