DeepInfer FAQ: DeepSeek V4.1-Flash Questions
Short, sourced answers about DeepInfer, DeepSeek's V4.1-Flash inference framework: what it is, what it runs on, what's missing, how fast it is, and which DeepInfer they mean.
The one-sentence answer: DeepInfer is DeepSeek's open-source inference framework for serving DeepSeek-V4.1-Flash at scale. It is an offline engine, not a hosted API, and it does not include the KV cache, prefill/decode disaggregation, or load balancing needed for a complete service. This FAQ covers the V4.1-Flash release, hardware, benchmarks, and name collisions.
DeepInfer supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, but DeepSeek's published performance charts cover NVIDIA hardware only.
Basics
What is DeepInfer?
DeepInfer is DeepSeek's open-source inference framework for serving DeepSeek-V4.1-Flash at scale. It was released on 20 October 2026 under the Apache-2.0 license.
Is DeepInfer a model?
No. DeepInfer is the engine that runs a model. The model is DeepSeek-V4.1-Flash-1017, whose weights are MIT-licensed.
Is DeepInfer free for commercial use?
The DeepInfer code is Apache-2.0, which permits commercial use and modification with the usual notice obligations. The default model weights are MIT. Check the licenses of bundled dependencies before deploying.
Running it
Which models does DeepInfer support?
Only DeepSeek-V4.1-Flash in this release. The default checkpoint is DeepSeek-V4.1-Flash-1017.
Can I run DeepInfer on an RTX 4090 or H100?
The README lists NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Older NVIDIA generations and consumer cards are not listed as supported.
Does DeepInfer include an OpenAI-compatible API server?
No. The release provides offline inference, not a hosted API. You would need to build the serving layer. DeepSeek also released deepseek-recipe, which handles API protocol conversion and tokenization, but it is not a server.
Is DeepInfer a complete production serving stack?
No. The README explicitly lists three missing pieces: a KV cache system, VE/Prefill/Decode disaggregation, and load balancing across data-parallel replicas.
The gap
Why is the KV cache system missing?
DeepSeek says the KV cache management and load-balancing components are tightly coupled to its internal infrastructure, so that code is not fully open-sourced yet.
Does DeepInfer support Huawei Ascend 950?
Yes, the README lists Ascend 950 NPUs alongside NVIDIA SM100 GPUs. DeepSeek did not publish Ascend 950 performance numbers for DeepInfer.
Did DeepSeek really open-source its production engine?
DeepSeek says DeepInfer powers its production services, but the release is a subset. The KV cache system, PD disaggregation, and load balancing are not included, so production parity cannot be verified from outside.
Numbers
How fast is DeepInfer?
DeepSeek published two Pareto charts: prefill on 4 SM100 GPUs and decode on 64 SM100 GPUs at a 131,072-token context. There is no third-party reproduction. Read the charts as trade-off maps, not as promises for your workload.
What is DSpark?
DSpark is DeepSeek's speculative-decoding path. It uses draft tokens and a confidence head to schedule variable-length verification. The default maximum draft length is 5.
What is SWA bounded replay?
It reconstructs missing sliding-window-attention KV by replaying only the most recent tokens instead of the full window across every layer. V4.1 defaults keep a 140-token suffix.
What is deepseek-recipe?
A separate MIT-licensed Rust/Python library that converts Chat Completions, Responses, and Messages requests into DeepSeek prompts and token IDs, and parses model output back into protocol responses.
Names
The name DeepInfer also refers to an unrelated 2016 medical-imaging toolkit and an unrelated 2023 smart-contract type-inference project; this site covers only DeepSeek's 2026 inference framework.
Is DeepInfer the same as DeepInfra?
No. DeepInfra is a commercial inference provider at deepinfra.com. DeepInfer is DeepSeek's open-source framework. DeepInfer is not affiliated with DeepInfra, the commercial inference provider.
Is DeepInfer the same as the medical or smart-contract DeepInfer?
No. The 2016 medical-imaging toolkit and the 2023 smart-contract research project are unrelated to DeepSeek's 2026 inference framework.
Where to start
If you want the release facts, read What DeepInfer actually ships. If you want the engineering, read Inside the engine. If you are deciding whether to use it, read the limits first, then the comparisons: DeepInfer vs vLLM, vs SGLang, and vs TensorRT-LLM. If you are confused by the name, read the disambiguation page.
Is DeepInfer better than vLLM?
There is no public head-to-head. vLLM has the broader serving layer, OpenAI-compatible API, and model coverage; DeepInfer is DeepSeek's one-model engine with DeepSeek-specific tuning. See the DeepInfer vs vLLM comparison.
Should I use DeepInfer or SGLang?
SGLang has a more complete V4.1-Flash preview path with an API, PD disaggregation, HiCache, and a router. DeepInfer has DeepSeek's own engine and Ascend 950 support. See the DeepInfer vs SGLang comparison.
We have not run DeepInfer. We do not have SM100 or Ascend 950 hardware. The answers above are source-based; where a claim is DeepSeek's own and not independently verified, we say so on the page that discusses it.