---
id: "2026-10-20-deepseek-open-sources-deepinfer"
url: "https://postcutoff.com/e/2026-10-20-deepseek-open-sources-deepinfer/"
as_of: "2026-10-20T15:06:00+02:00"
date: "2026-10-20"
date_precision: day
category: open-source
importance: 3
confidence: high
status: [Confirmed]
sources: 1
editor: human editor
human_review: null
version: "2026-10-20"
---

As of: 2026-10-20 15:06 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-20-deepseek-open-sources-deepinfer/

# DeepSeek open-sources DeepInfer, the inference framework behind its production services, with Nvidia SM100 and Huawei Ascend 950 support

DeepSeek published DeepInfer on GitHub under Apache 2.0, describing it as the high-performance inference framework that powers its large-scale production services. The release covers offline inference for DeepSeek-V4.1-Flash on Nvidia SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs; the KV-cache system, prefill/decode disaggregation and cross-replica load balancing are not included.

## Key facts
- Repo: github.com/deepseek-ai/DeepInfer, Apache-2.0; ~4.9k stars within days, 328 points on Hacker News (Oct 20)
- Supported model: DeepSeek-V4.1-Flash (default checkpoint DeepSeek-V4.1-Flash-1017)
- Hardware: Nvidia SM100 GPUs (CUDA 13) and Huawei Ascend 950 NPUs (CANN 9.2)
- Features: optimized kernels; attention data parallelism, balanced context parallelism, MoE expert parallelism with EPLB; 'DSpark' decoding with draft tokens and confidence-scheduled verification; asynchronous zero-overhead batch scheduling; batch-invariant deterministic sampling (Nvidia only); structured outputs via regex, JSON schema or EBNF
- Benchmarks are given only as Pareto-curve images (setup: 4 SM100 GPUs for prefill, 64 for decode)
- Not included: KV cache system, prefill/decode disaggregation, load balancing across data-parallel replicas

## What happened

DeepSeek released the production inference stack for its V4.1 Flash model, continuing its habit of open-sourcing infrastructure.
First-class support for Huawei's Ascend 950 alongside Nvidia Blackwell shows the stack is built for both hardware lines.

## Why it matters

Serving efficiency is a large part of DeepSeek's low prices; publishing the engine lets others reproduce it, and Ascend support
matters as China pushes domestic accelerators. The release came the day after the US banned DeepSeek from federal civilian agencies.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 173 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 112 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 203 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 142 days after its cutoff

## Sources

1. [GitHub: deepseek-ai/DeepInfer](https://github.com/deepseek-ai/DeepInfer) (github.com, code)

## Changes

- 2026-10-20 (filed): Created (GitHub watch + Hacker News)

## Related

- 2026-10-19: [DHS orders all US federal civilian agencies to remove DeepSeek products](https://postcutoff.com/e/2026-10-19-dhs-bans-deepseek-federal-civilian-agencies/index.md)
- 2026-10-17: [DeepSeek-V4.1-Flash-1017 update: big agentic-coding gains, same price](https://postcutoff.com/e/2026-10-17-deepseek-v4-1-flash-1017/index.md)
- 2026-09-29: [DeepSeek partners with Huawei, open-sourcing TileLang-based programming tools and libraries for Ascend chips as a CUDA alternative](https://postcutoff.com/e/2026-09-29-deepseek-huawei-tilelang-ascend/index.md)