MiMo-V2.6-Flash
Released Sep 21, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 128K
- Speed
- —
- Price ($/MTok in / out)
- $0.14 / $0.28
- Cost per task
- —
- Open weights
- Yes
Benchmarks1 of 10 reported
Terminal-Bench 2.1
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Stays within a few points of Pro on Xiaomi's agent suite while activating about a third of the parameters — it was trained in the same RL run, so the agent behaviour is the same family, just cheaper to serve
- Small enough to run outside a data centre — a GGUF conversion and DGX Spark ports appeared within days, and OpenCode gave it away free for launch week
- Edges Pro at reproducing real-world vulnerabilities — the one place in the family table where the smaller model leads
- Same omnimodal input set, 1M context and thinking on/off switch as Pro, so a pipeline built on Pro can drop to Flash without changing the request shape
Weaknesses
- Runs away in agent harnesses — the first Hugging Face reports on the card describe it issuing thousands of repeated searches in OpenCode and burning tokens on useless work without finishing
- Declares a job done when its own tests pass — a hands-on OpenCode run had it drop a required subcommand from a multi-file feature that its self-written tests never exercised, so check output against the brief rather than its tests
- Offensive security is what shrinks most going down from Pro — its exploit-writing scores fall to roughly a third to a half of Pro's while vulnerability reproduction stays level
- Too new to characterise beyond launch week — Artificial Analysis has not yet run it, so its speed, per-task cost and independent knowledge scores are unmeasured
The efficiency checkpoint of Xiaomi's V2.6 series: a 309B-parameter sparse MoE activating 15B, natively omnimodal (text, image, video and audio in; text out), trained in the same livestreamed 30-step RL run as Pro at a cost Xiaomi puts at $0.85M. Announced 2026-09-21 UTC (22 September Beijing time on mimo.mi.com). Terminal-Bench 2.1 87.6 is Xiaomi's own figure from the Hugging Face card and tech report Table 3. Also reported by Xiaomi: DeepSWE v1.1 67.9, ProgramBench 26.0, AutomationBench v1.0.6 52.3, Toolathlon-Verified 73.6, Agents' Last Exam 27.6, Terminal-Bench 4.0 28.8, OSWorld-Verified 80.8, JobBench 61.2, CyberGym 95.1, ExploitGym 6.0, ExploitBench 25.3, SEC Bench Pro 47.5 and MiMo Visual Coding 71.5; Xiaomi's table has no GDPval-AA entry for Flash. GPQA Diamond, SWE-bench Verified/Pro, HLE, AIME, MMLU-Pro and ARC-AGI-2 are unpublished, it is not on the arena.ai text board (last update 2026-09-13), and Artificial Analysis has no Flash listing as of 2026-09-24, so costPerTask and speed are null. Pricing is Xiaomi's pay-as-you-go rate ($0.14 cache-miss input, $0.0028 cache-hit, $0.28 output, unchanged from V2.5). Context 1M and max output 128K per the mimo.mi.com model page. Weights (MiMo-V2.6-Flash-RL) are MIT on Hugging Face and ModelScope, declared in the card metadata with no separate LICENSE file; a ggml-org GGUF exists. Knowledge cutoff is unpublished — the "December 2024" string on the model page is a sample system prompt. predecessorId is null: no Xiaomi source says it replaces MiMo-V2.5 or V2-Flash, neither of which is on this site.
For developers
API model strings
- xiaomi
mimo-v2.6-flash - openrouter
xiaomi/mimo-v2.6-flash
Licence
MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor