• Bottom line: Wait. The repository is offline inference only, so it is not a drop-in production serving stack.
• The README lists only one model, DeepSeek-V4.1-Flash, as supported, on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.
• A KV cache system, prefill/decode disaggregation and load balancing across replicas are not included; the README says a complete service needs them.
• Neither the Decoder report nor the README text we could read gives a cost or energy figure, so the savings claim is unmeasured.
• The Apache-2.0 license and a reproduction script (examples/sweep/run_sweep.sh) let teams check the throughput charts themselves.

Wait. DeepSeek has open-sourced DeepInfer, the inference framework behind its V4.1-Flash model, under the Apache License 2.0, according to The Decoder. The repository is a real engineering artifact, but the README limits it to offline inference, one model and two hardware families. The headline claim of cheaper, greener inference is DeepSeek's stated goal, not a number either source we read puts in front of the reader.

What was actually released

The README calls DeepInfer the high-performance framework that powers DeepSeek's large-scale production services, but says this release covers offline inference only. It is built primarily on vLLM and adds optimized kernels such as FlashMLA, DeepGEMM and DeepEP, attention data parallelism, expert parallelism with load balancing, and speculative decoding. The Decoder also reports a companion tool, deepseek-recipe, meant to make self-hosted token processing match the official API.

Where it falls short of a serving stack

The README states that a KV cache system, prefill/decode disaggregation and load balancing across data-parallel replicas are not included, and that a complete service requires them. For an enterprise team, that means the repository supplies the engine core, not the surrounding production system. Anyone planning to serve users would still have to build or source those components before comparing the result with a hosted API.

Hardware and model constraints

The README lists DeepSeek-V4.1-Flash as the only supported model and names NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs as hardware. It requires Linux, Python 3.12 and GCC 13 or newer, plus CUDA 13 or CANN 9.2. Its deterministic, batch-invariant sampling is NVIDIA-only. Teams on older GPUs, or running other models, cannot use the release as published.

The cost claim is unmeasured

DeepSeek said on X its aim is to help the community lower compute costs, reduce energy use and make AI more affordable. The Decoder gives no price or energy figures, and the README text we could read shows prefill and decode Pareto charts without specific numbers. The prefill runs used 4 SM100 GPUs and the decode runs 64. That is a claim to test, not a result to cite.

What to watch next

The Decoder reports that code for future DeepSeek models will also ship as open source, and frames the release as added pricing pressure on Western providers such as OpenAI and Anthropic. The practical test is independent reproduction: run examples/sweep/run_sweep.sh on comparable hardware and compare throughput per GPU with an existing vLLM deployment. Until others publish those numbers, treat the efficiency story as unverified.

The bottom line

The bottom line: Wait. The number that decides it is one, the count of supported models, alongside offline-only inference and SM100 or Ascend 950 hardware, with no KV cache or load-balancing layer. Read the code if you operate DeepSeek models on supported hardware, and reproduce the sweep before trusting any savings. Everyone else should hold until a complete serving layer ships or independent benchmarks show a measurable cost advantage over current vLLM setups.