MiMo-V2.6-Pro
Released Sep 21, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 128K
- Speed
- 51 tok/s · 2.83s to first answer token
- Price ($/MTok in / out)
- $0.43 / $0.87
- Cost per task
- $0.13
- Open weights
- Yes
Benchmarks2 of 10 reported
Terminal-Bench 2.1
Humanity's Last Exam
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
AIME (latest)retired
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Coding skill carries between scaffolds rather than living in one harness — Xiaomi trained across four mini-harnesses and its held-out Codex, Claude Code and mini-swe-agent runs improved in step, so it needs less re-prompting when moved between tools
- One checkpoint takes text, images, video and audio, so screenshot-driven computer use and visual coding go through the same weights as the terminal work — no separate vision endpoint to route to
- Cache reads are priced at effectively nothing, so an agent loop that re-reads a large context every turn costs little more than its fresh tokens — and the whole RL run was livestreamed with the log left public, with the environments and framework being released alongside the weights
- Finishes small, well-specified coding jobs unaided — a hands-on OpenCode run had it complete all three tasks and pass the same hidden checks as Claude Opus 5 at a small fraction of the spend
Weaknesses
- Can fall into repeated tool-call loops in agent harnesses — early Hugging Face reports on the Pro and Flash cards describe it cycling through the same searches hundreds of times in OpenCode without finishing
- Guesses rather than declines when it does not know — Artificial Analysis finds it answers wrongly on roughly four in ten of the knowledge questions it misses, and its knowledge score sits near the bottom of the index
- Thinks at length for its price class — Artificial Analysis calls it somewhat verbose and a single Intelligence Index task takes it around eighteen minutes end to end, so the low per-token rate is partly spent on tokens
- Finds and reproduces vulnerabilities far better than it writes exploits — its own table puts it level with the closed frontier on CyberGym but at well under two-thirds of them on ExploitBench
Flagship of Xiaomi's V2.6 series: a 1.02T-parameter sparse MoE activating 42B per token, natively omnimodal (text, image, video and audio in; text out), trained with a 30-step, livestreamed RL run Xiaomi costs at $2.62M. Announced 2026-09-21 UTC (Xiaomi's tweet and Artificial Analysis; mimo.mi.com's news page dates it 22 September, Beijing time). Terminal-Bench 2.1 89.9 is Xiaomi's own figure from the Hugging Face card and tech report Table 3, whose baselines run at max effort; HLE 49.4% is Artificial Analysis' independent text-only, no-tools measurement, since Xiaomi publishes none. GDPval-AA is 1673 on v2.1 (Xiaomi's table and AA agree), a different scale from the v2 this key records, so that cell is null. Also reported by Xiaomi: DeepSWE v1.1 71.9, ProgramBench 26.5, AutomationBench v1.0.6 53.1, Toolathlon-Verified 76.9, Agents' Last Exam 31.6, Terminal-Bench 4.0 34.9 (AA measures 34.8), OSWorld-Verified 82.0, JobBench 62.0, CyberGym 94.0, ExploitGym 17.8, ExploitBench 47.9, SEC Bench Pro 66.3 and MiMo Visual Coding 72.3. GPQA Diamond, SWE-bench Verified/Pro, AIME, MMLU-Pro and ARC-AGI-2 are unpublished for Pro and Flash (the report's SWE-bench rows are for the 9B distillation only), and neither is on the arena.ai text board (last update 2026-09-13). AA scores it 46 on Intelligence Index v4.3.2, first among open-weights models; costPerTask ($0.1332) and speed (51.3 tok/s, 2.83 s to first token as AA headlines it) are AA's measurements of its single reasoning variant — the API exposes a thinking on/off switch and no effort levels, so effort is null; launch coverage quoted AA at about 130 tok/s before the page settled. Pricing is Xiaomi's pay-as-you-go rate ($0.435 cache-miss input, $0.0036 cache-hit, $0.87 output, unchanged from V2.5); a Pro-UltraSpeed endpoint at ten times the price ($4.35/$8.70) promises up to 20x faster output and is not a separate entry. Context 1M and max output 128K per the mimo.mi.com model page (OpenRouter lists 1,048,576 and 131K). Weights (MiMo-V2.6-Pro-RL) are MIT on Hugging Face and ModelScope, declared in the card metadata with no separate LICENSE file. Knowledge cutoff is unpublished — the "December 2024" string on the model page is a sample system prompt, not a spec. predecessorId is null: no Xiaomi source says V2.6-Pro replaces V2.5-Pro (the V2.5 suite stays on sale) and MiMo-V2.5 is not on this site.
For developers
API model strings
- xiaomi
mimo-v2.6-pro - openrouter
xiaomi/mimo-v2.6-pro
Licence
MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor