Xiaomi MiMo-v2.6-Pro hands-on: agentic coding score triples, 900tps peak speed
karminski3 · x · 2026-09-23
A hands-on review of Xiaomi's MiMo-v2.6-Pro. Agentic coding jumped: the vector-database benchmark score tripled from 2505 (v2.5-Pro) to 7810, nearing Claude Fable-5, with the algorithm upgraded to single-graph HNSW (M=16/M0=28) plus AVX-512 exact distance and per-thread visited stamps.
Impressive speed: the ultraspeed variant hit 464tps in testing; the tech report claims 900tps peak and 500tps sustained — notable given the 1.02T total / 42B active parameter scale, where peers usually run 60-80tps.
Weaknesses: front-end ability lags (failed a ball-breaking-wall physics demo due to poor spatial understanding); early-stopping issues (gave up around iteration 30 of 50 in two of three runs); long thinking with no intensity control; the standard API is congested, so set generous timeouts. Verdict: strong for back-end agentic coding, weak on front-end — consistent with 67% coding-heavy post-training data from its livestreamed RL.
Related event: Xiaomi's MiMo-V2.6 Pro tested: big gains in coding, 3D games and multimodal(2 posts)→
More from coding & agent
- Anthropic: Opus 5.5 is cheaper, more token-efficient, and spans all effort levels — trq212 · 2026-09-23
- Hands-on With Opus 5.5: Same Feel as 4.6, More Power, ~40% Cheaper Than Opus 5 — EricBuess · 2026-09-23
- One-Paragraph Prompt Matches Your 'Carefully Designed' Autoresearch Pipeline — generativist · 2026-09-23
- Auto-research frameworks may be overkill: one paragraph prompt gets similar results — generativist · 2026-09-23
- Why Nautilo built its own mobile apps instead of piggybacking on chat apps — Dan_Jeffries1 · 2026-09-23
- One prompt turns Grok into a full marketing team: ads, videos and influencer outreach — JaynitMakwana · 2026-09-23