THE HOOK
DeepSeek has stopped just giving away models and started giving away the machine that serves them.
THE GIST
DeepSeek released DeepInfer on GitHub under Apache-2.0, describing it as the framework behind its large-scale production services. It runs DeepSeek-V4.1-Flash for offline inference on Nvidia SM100 GPUs and Huawei Ascend 950 chips. It's free, but the KV cache system, prefill/decode split and load balancing aren't included.
THE TAKE
This is the most useful thing DeepSeek has shipped in months, and it isn't a model. Inference cost is where the real margins live, and now everyone can read the recipe. The catch is the hardware list: Blackwell-class Nvidia or Huawei's newest NPUs, nothing in between. Open source, closed garage.
OPINION
The Interconnects roundup covered by AI Weekly cites estimates that DeepSeek makes up to 90% gross margin selling tokens on this stack, which explains why its prices keep embarrassing rivals. Landing the same week Washington barred DeepSeek from federal procurement, it's also a neat reminder that you can ban a vendor but not an Apache licence.
- Story type: Important story
- Impact: Medium
- Confidence: High (official announcement)
- Audience: Developers
- What changed: DeepSeek open-sourced DeepInfer, its production inference framework, under Apache-2.0.
- What to watch: Whether DeepSeek releases the missing KV cache and disaggregation pieces, and how fast vLLM and SGLang absorb its kernels.
- Reading time: 1 minute
