Huawei Atlas 950 lineup details leak as enthusiast runs GLM and Muse on Atlas 300I Duo cards
matteiuspi · reddit · 2026-10-22
A long-time Huawei Atlas tinkerer shared fresh first-hand info on Reddit:
- New 950/150 series intel: obtained via a Silver-rated Huawei partner; the info was Chinese-only so he translated it into an English chart. The brand-new Atlas 150 uses 950 Ascend cores, and he's working on purchase options.
- Hands-on progress: running 3x passive-cooled Atlas 300I Duo (96GB) cards, he keeps optimizing QFN in his vLLM fork, testing GLM5.3-flash and DeepSeek models; this week he ran Meta Muse after a REAP expert-pruning pass to fit the 3-card setup — still too slow.
- Use cases: Qwen video models describing security camera feeds; a friend generating images on the same cards.
- Practical notes: needs custom fan shrouds (passive cooling), an extra PCIe 12V CPU power cable, solid Linux skills, and the embedded Euler OS firmware signature verification can't be bypassed yet.
He argues Huawei is China's strongest competitor to NVIDIA and that this competition will normalize ML/LLM pricing globally.
More from coding & agent
- NVIDIA open-sources SkillSpector after scan finds 26% of 31k agent skills vulnerable — dunkhippo33 · 2026-10-22
- AgentMail hits 1B emails processed, up 50x in a year, as agents need inboxes — rohanpaul_ai · 2026-10-22
- UCLA x Apple open-source OpenVLHarness multimodal agent, gaining 9.7-12.6 points across 22 benchmarks — WeijiaShi2 · 2026-10-22
- One-shot RL repro: training Qwen3.5-4B to paint watercolour flowers with a judge model — charles_irl · 2026-10-22
- Open-source Claude Duet alternative Open-Duet hits 5.9k stars on launch day — matchaman11 · 2026-10-22
- LGWM: A World Model That Verifies GUI Agent Transitions and Detects Hijacked Screens — Jiaming Zhang · 2026-10-22