Radar / AI infrastructure / DeepSeek open-sources Ascend infrastructure…
DeepSeek open-sources Ascend infrastructure stack
DeepSeek published Ascend-optimized versions of its NVIDIA-proven infra components — TileLang Ascend, DeepGEMM-Ascend, DeepEP-Ascend, FlashMLA Ascend, TileKernels Ascend backend, and DeepSelect — developed with Huawei's support around the Ascend 950 128-card supernode.
Why it matters
Ports the exact kernel stack that trained DeepSeek V4 (TileLang carrying most V4 training operators, DeepGEMM near hardware peak) 1:1 to Huawei Ascend NPUs, so any team can train and serve MoE models on domestic Chinese chips without rewriting Ascend C kernels by hand — a direct strike at CUDA lock-in.
What you could build with it
An infrastructure team or chip startup can reuse the TileLang-Ascend compiler plus DeepEP-Ascend for expert-parallel communication to stand up an Ascend-native training or serving stack for their own MoE model, instead of hand-writing Ascend C operators and NCCL-equivalent collectives.
Does it hold up?
Too early to judge: announced 2026-09-30, and third-party adoption evidence is limited to press writeups so far. The performance claims (95% of peak prefill, 90-95% of physical bandwidth for DeepEP) come from DeepSeek's own docs, and some DeepEP ops remain experimental with manual config not yet publicly distributed.
Built with DeepSeek open-sources Ascend infrastructure…
- 网易智能：DeepSeek这次开源，想让华为芯片更好用article · In-depth writeup of the announcement: every TileLang operator used in V4 training now has an Ascend implementation; FlashMLA hits 95% of hardware peak in prefill and 83% in decode on Ascend 950; Huawei co-developed the stack and is pushing a 128-card supernode.
- 凤凰网科技：DeepSeek开源昇腾基础组件，与英伟达平台一一对应article · Coverage of the 9/30 announcement listing the components (DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect, TileLang Ascend) and noting DeepSeek's strategic push to train on domestic chips.
- 联合早报：DeepSeek开源面向华为昇腾平台基础设施组件article · Singaporean coverage noting the stack mirrors DeepSeek's earlier NVIDIA opens and quoting Liang Wenfeng's estimates of ~200K Ascend 950 chips needed to match ~50K NVIDIA GB300s for frontier training.
- Reuters: DeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidiaarticle · Reuters' report on the 9/30 announcement: DeepSeek partnered with Huawei to open-source programming infrastructure for Ascend NPUs (announced on DeepSeek's official WeChat), including compute and communication libraries, plus a 128-card Ascend 950 supernode solution.
- tile-ai/tilelanggithub · 7,865 ★ · TileLang's own repo (now with the Ascend backend DeepSeek used for most V4 training operators); unified Python API picks the GPU/NPU backend at runtime.
- deepseek-ai/DeepEP-Ascendgithub · 173 ★ · Ascend port of DeepEP for large-scale expert-parallel dispatch/all-to-all communication, with the public API aligned to the NVIDIA version.
Learn more
First spotted on github: source.
More AI infrastructure
Modular open-sources MAX inference server, Mojo stdlib and accelerator kernelsThe core of Modular's unified AI deployment stack is now open: the MAX inference server (OpenAI-compatible endpoints)…infra · JEV 0.76OmniRouteFree MIT AI gateway: one endpoint, 290+ providers and 500+ models with quota-aware auto-fallback.infra · JEV 0.75Cloudflare Agents Week: Sandboxes GA, 50K concurrent Workflows, Managed OAuth for agentsA dozen agent-infrastructure launches in one week: persistent Linux Sandboxes (GA) with real shell/filesystem/state…infra · JEV 0.69ai-memory: long-term memory for agent coding CLIsRust solution for long-term memory for agent coding CLIs, facilitating handoff between different agents and sessions.infra · JEV 0.68Hindsight: agent memory that learnsAgent memory system built for learning over time: retain/recall/reflect operations with SOTA scores on the LongMemEval…infra · JEV 0.63DeepSeek Elastic Compute (DSec): Sandbox Infrastructure for Agentic Training at ScaleDeepSeek's production sandbox platform behind all its RL training from V3.2 through V4.1: a unified SDK over FnCall…infra · JEV 0.6
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.