Xiaomi MiMo-V2.6 Review 2026: 1.02T Parameters, 46 on AA's Open-Weights Index — Trained in 6 Days for $3.47M
Xiaomi MiMo-V2.6 Review 2026: 1.02T Parameters, 46 on AA's Open-Weights Index — Trained in 6 Days for $3.47M
Verdict first: Xiaomi MiMo-V2.6 is the rare open-model release where the headline number is not the vendor's. MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index v4.3.2 — measured by Artificial Analysis with its own harness, not self-reported by Xiaomi — making it the highest-ranked open-weights model in the world, ahead of GLM-5.3 (45) and Kimi K3 (44), and a startling 20 points above its own predecessor MiMo-V2.5-Pro (26). The model is a sparse MoE with 1.02 trillion total parameters and 42B active per token, a 1M-token context, and native text, image, video and audio input, under an MIT license. The economics are the other half of the story: $0.43 in / $0.87 out per million tokens with a 99% cache discount, which Artificial Analysis measured at $0.13 per index task — frontier-adjacent intelligence for about a dime. Xiaomi got there by scaling reinforcement learning rather than pre-training: same architecture, same parameter count, under 6 days and roughly $3.47M. The honest caveats: it is 5 points behind Claude Fable 5.1 and GPT-6 Astra (both 53), its headline DeepSWE gains have not been submitted to a public leaderboard, training costs are self-reported, per-benchmark breakdowns are unpublished, and there is a single serving route with no failover.
Xiaomi MiMo-V2.6-Pro — 1.02T total / 42B active, 1M context, omnimodal input, MIT licensed: open-weights leader at 46, five points off the closed frontier
What Is MiMo-V2.6?
MiMo-V2.6 is Xiaomi's new flagship model family, released and open-sourced in the early hours of September 22, 2026. The interesting claim is not that it is bigger. It is that Xiaomi says it got there without making the base model bigger.
The company's framing: by scaling reinforcement learning compute while holding pre-training architecture and parameter count fixed, capability jumped. If that path generalises, it matters more than another parameter race — pre-training at this scale is affordable to only a handful of labs, and post-training that produces this kind of gain is a far cheaper lever.
The most direct evidence for the claim is generational. MiMo-V2.5-Pro scored 26 on the Artificial Analysis Intelligence Index. MiMo-V2.6-Pro scored 46. A twenty-point jump, with the base model unchanged, is the number that makes this release worth reading closely.
Variants and Specs
| Variant | Positioning | Parameters |
|---|---|---|
| MiMo-V2.6-Pro | Flagship, natively omnimodal | Sparse MoE, 1.02T total / 42B active |
| MiMo-V2.6-Flash | Lightweight tier | Sparse MoE, 309B total / 15B active |
| MiMo-V2.6-Pro-UltraSpeed | High-speed hosted tier | Same checkpoint, up to 20× output speed claimed (third-party listings say ~10×) |
| MiMo-V2.6-Distill-Qwen-9B | Distilled model for community RL research | 9B |
Shared specifications:
- Context: 1M tokens — the docs describe 1M-context training; the inference-side ceiling is not stated numerically
- Modalities: native text, image, video and audio input — Pro and Flash are both natively omnimodal, and Xiaomi describes Pro as the only omnimodal Pro model in the open-weights camp
- Model IDs (must be lowercase):
mimo-v2.6-pro,mimo-v2.6-flash,mimo-v2.6-pro-ultraspeed - License: MIT
A MiMo Desktop client also left beta alongside the models, with membership plans.
The 46: Why This Score Is Different
| Model | AA Intelligence Index v4.3.2 | Notes |
|---|---|---|
| Claude Fable 5.1 | 53 | Frontier (max tier) |
| GPT-6 Astra | 53 | Frontier (max tier) |
| Claude Opus 5 | 51 | Closed |
| Xiaomi MiMo-V2.6-Pro | 46 | Open-weights leader |
| GLM-5.3 | 45 | max tier |
| Qwen3.8 Max | 45 | — |
| Kimi K3 | 44 | max tier |
| MiMo-V2.5-Pro (predecessor) | 26 | — |
The detail that matters most: the 46 was measured by Artificial Analysis using its own harness, not reported by Xiaomi. That is genuinely uncommon for a Chinese model launch, where benchmark figures are near-universally vendor-run. The index spans ten evaluations across reasoning, knowledge, math and coding — AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1 — run by a party with no stake in the result.
What the score does and does not mean:
- ✅ Top open-weights model — ahead of GLM-5.3 (45), Qwen3.8 Max (45) and Kimi K3 (44)
- ❌ Not the top overall — it sits 5 points behind Claude Fable 5.1 and GPT-6 Astra (both 53), with Claude Opus 5 at 51 above it as well
- ⚠️ Hugging Face CEO Clément Delangue reshared the result with "Banger!" — but a reshare is not a reproduction; the independent evidence remains AA's own run
Serving characteristics (measured by AA):
| Metric | Value | Context |
|---|---|---|
| Output speed | 134.3 tokens/sec | 11th of 114 models (median 77.9) |
| Time to first token | 2.15 s | Median 2.34 s |
| Cost per index task | $0.13 | First sub-$0.20 frontier-adjacent result |
| Full evaluation run | $206.66 | — |
Pricing: Same as Last Generation, Better Model
| Item | Price (per 1M tokens) |
|---|---|
| Input | $0.43 (¥3) |
| Output | $0.87 (¥6) |
| Cache discount | 99% |
| Blended (7:2:1 cache-hit/input/output) | $0.18 |
| Flash tier | ¥1 in / ¥2 out |
Xiaomi reused the MiMo-V2.5 price sheet rather than raising it for the new checkpoint — the capability jumped 20 index points and the price did not move. That is the most concrete thing in this release.
Put it against the frontier: Claude Opus 5 is priced at $5.00 input / $25.00 output and scores 51 — five points higher, at roughly an order of magnitude more expensive per token. Xiaomi claims parity-intelligence pricing of 1/20th to 1/60th of overseas models. Note that UltraSpeed is a separate, notably higher-rate tier for latency-sensitive work, not the default.
Inside the RL Scaling Recipe
Xiaomi disclosed more methodology than most labs do, which makes the claim checkable in principle.
Compute scaled along three axes:
- Larger batch and throughput — 1,568 samples per update, 1M context support, 2.7–3.7B tokens per step
- More tasks and harder environments — Code, General, Visual and Cyber task families, with multiple mixed harnesses
- More grader compute — expanded scoring-side capacity
The training bill:
- ~$2.62M for Pro, ~$854K for Flash — about $3.47M combined (self-reported, unaudited)
- 30 RL steps each, roughly 750,000 trajectories, completed in under six days
- Average task pass rate improved 25% (Pro) and 12% (Flash)
Stability work — worth calling out, because this is where large-scale RL usually breaks: Xiaomi froze the MoE router to suppress expert load drift, and built explicit reward-hacking defenses.
What is not verified: the DeepSWE v1.1 gains Xiaomi reports (58.4 → 72.57) were run with mini-swe-agent, averaged over three attempts, and not submitted to a public leaderboard or independently reproduced. The public leaderboard's top tier sits around 74% Pass@1 under submitted configurations — a different measurement, so the two are not directly comparable. Treat the dashboard numbers as directionally encouraging rather than settled.
What Xiaomi Actually Open-Sourced
This is the most underrated part of the release. Xiaomi shipped considerably more than weights.
| Released | Detail |
|---|---|
| Pro / Flash weights | MIT license |
| Technical report | — |
| 7,000+ high-quality RL task environments | Software engineering, vulnerability reproduction, knowledge work, web design & development |
| End-to-end RL training framework | Built on verl, uni-agent and mini-swe-agent |
| Lightweight composable mini-harnesses | — |
| MiMo-V2.6-Distill-Qwen-9B | A starting point for community RL research |
7,000+ RL task environments plus an end-to-end training framework is worth substantially more than weights alone — it lets others reproduce the post-training path, not just run inference against the result.
Evidence from the distilled model — Distill-Qwen-9B improved over its SFT baseline on all 11 evaluations, which is the cleanest demonstration of what the RL recipe contributes:
| Benchmark | SFT baseline | After RL |
|---|---|---|
| SWE-bench Verified | 61.1 | 66.2 |
| MiMo Cyber Bench | 31.3 | 47.0 |
| Terminal Bench 2.1 | 37.1 | 52.8 |
| MiMo Visual Coding | 64.0 | 72.4 |
The Honest Take
Score: 8.5/10 for the open-weights category. As a bid for the absolute capability crown it is not the point of this release, and it should not be graded as one.
What genuinely impresses:
- The score is independently measured. Artificial Analysis ran it, not Xiaomi. After years of writing "all benchmarks are vendor-reported" in open-model reviews, this is the single most refreshing thing about the release — and it makes the 46 meaningfully more useful than a self-reported 46 would be.
- A 20-point generational jump with the base model unchanged. 26 → 46 without touching pre-training architecture or parameter count is a strong result for the RL-scaling thesis.
- $0.13 per task. Frontier-adjacent intelligence at roughly a dime, measured by a third party.
- The price did not move. Same rate card as V2.5 for a substantially better model.
- 1.02T total with only 42B active, plus 1M context — the sparsity ratio is what keeps long-context, large-codebase workloads economically viable.
- Natively omnimodal across text, image, video and audio.
- The training accounting is disclosed — $2.62M / $854K, 30 steps, 750K trajectories, under six days. Publishing the bill is a form of confidence.
- 7,000+ RL environments and a full training framework make this reproducible in a way that "weights only" releases never are.
- Breadth of demonstrated work: 3D world reasoning, Blender modelling, embodied control (Franka Panda arm), computer-use agent, plus front-end, Figma, PPT, SVG, video and music generation — including a self-arranged orchestral piece for roughly ten instruments with MIDI export. In research, it assisted PFAS adsorbent material screening (compressing a month of work into 2–3 days) and completed a Lean 4 formalisation of the Li–Yorke theorem.
- MIT license — commercially permissive.
What should temper your excitement:
- Open-weights leader is not overall leader. 53, 53 and 51 sit above it.
- The DeepSWE gains are not on a public leaderboard. Vendor-run with mini-swe-agent over three attempts, against a public top tier measured differently. That is a real evidentiary gap.
- Training cost is self-reported and unaudited — neither an API price nor a third-party reconciled figure.
- Xiaomi has disclosed its own training failures: GPU OOM restarts from expert load imbalance, network faults between the training cluster and grader deployment, and a Flash restart after an infrastructure error that went unnoticed for about three hours. Credit for disclosing them; they are still real reliability signals.
- Per-benchmark breakdowns are unpublished. Which of the ten evaluations it wins or loses determines deployment fit, and a composite 46 does not tell you.
- A single serving route with no failover. For production that is a resilience problem, not a preference.
- Research and formal-math results are described as potential shown without targeted RL training — the materials case still needs wet-lab validation.
- UltraSpeed is a separate premium tier, not something you get by default.
Who should use it: teams that need frontier-adjacent capability on a tight budget — $0.13/task and a $0.18 blended rate make it one of the strongest price-performance positions available; anyone with 1M-context, long-horizon agent or large-codebase workloads; applications needing native image, video and audio input; and research groups wanting to do RL post-training, because 7,000+ task environments and a full training framework is a ready-made starting point. The MIT license makes it viable for commercial work.
Who should skip it: anyone needing the absolute capability ceiling; production systems requiring multi-route failover; teams that need published per-benchmark detail to make a selection decision; and procurement processes requiring audited cost figures.
FAQ
What is Xiaomi MiMo-V2.6?
Xiaomi's new flagship model family, released and open-sourced on September 22, 2026. MiMo-V2.6-Pro is a sparse MoE with 1.02T total and 42B active parameters, a 1M-token context window, and native text, image, video and audio input. MiMo-V2.6-Flash is the lighter tier at 309B total / 15B active. Both are MIT licensed. Pro scored 46 on the Artificial Analysis Intelligence Index, making it the highest-ranked open-weights model.
Is the 46 score vendor-reported?
No — and that is the most important thing about this release. It was measured by Artificial Analysis using its own evaluation harness, not reported by Xiaomi. That is unusual for a Chinese model launch, where figures are almost always vendor-run. The index covers ten evaluations across reasoning, knowledge, math and coding. It is the top open-weights score, not the top overall: Claude Fable 5.1 and GPT-6 Astra both score 53, and Claude Opus 5 scores 51.
How much does MiMo-V2.6 cost?
Xiaomi kept the MiMo-V2.5 price sheet: $0.43 per million input tokens and $0.87 per million output tokens, with a 99% cache discount and a $0.18 blended rate at a 7:2:1 cache-hit/input/output mix. Artificial Analysis measured $0.13 per index task, with a full evaluation run at $206.66. Flash tier pricing is ¥1 input / ¥2 output per million tokens. Xiaomi claims 1/20th to 1/60th of overseas model pricing at equivalent intelligence. UltraSpeed is a separate, higher-rate tier.
What does "RL scaling" mean here?
Xiaomi says it improved MiMo-V2.6 by scaling reinforcement learning while keeping the pre-training architecture and parameter count unchanged — the base model did not change, the post-training did the work. Compute scaled three ways: larger batch and throughput (1,568 samples per update, 1M context, 2.7–3.7B tokens per step), more tasks and environments (Code, General, Visual, Cyber with mixed harnesses), and more grader compute. Cost was roughly $2.62M for Pro and $854K for Flash — about $3.47M combined — over 30 steps each, around 750,000 trajectories, in under six days.
What can it do that is unusual?
Native omnimodal input across text, image, video and audio; 3D spatial reasoning; computer-use agent capability; content creation spanning front-end code, Figma, PPT, SVG, video and music — including a self-arranged orchestral piece for roughly ten instruments with MIDI conversion. In research it assisted PFAS adsorbent screening, compressing a month of work into 2–3 days, and completed a Lean 4 formalisation of the Li–Yorke theorem. Note that the research and formal-math results are described as potential shown without targeted RL training, and the materials case still requires wet-lab validation.
What are the catch points?
Several. Its reported DeepSWE improvement (58.4 → 72.57) was run with mini-swe-agent averaged over three attempts, not submitted to a public leaderboard, and not independently reproduced — the public top tier is a differently-measured 74% Pass@1. Training costs are self-reported and unaudited. Xiaomi has disclosed its own training incidents, including GPU OOM restarts, cluster-to-grader network faults, and a Flash restart that went unnoticed for about three hours. Per-benchmark breakdowns are unpublished. There is a single serving route with no failover. And UltraSpeed is a premium tier rather than default behaviour.
Sources
- Xiaomi MiMo official docs — Xiaomi MiMo-V2.6 Series: 3 New Models Officially Released
- OrcaRouter — Xiaomi MiMo-V2.6-Pro: scored 46 on the Open-Weights Index (AA measurement data, pricing, limitations)
- Jiemian News — Xiaomi releases and open-sources the MiMo-V2.6 series (Sep 22, 2026)
- ITHome — MiMo-V2.6 released and open-sourced: Pro and Flash versions, API pricing unchanged
- 21st Century Business Herald — Xiaomi tops the open-source leaderboard: six days of training cost $3.47 million
- The Cover — Xiaomi MiMo-V2.6 tops the AA global open-source rankings
- CNMO — Xiaomi officially releases and open-sources the new MiMo generation
- Chinaz — Xiaomi open-sources MiMo-V2.6: AA intelligence index 46 tops the strongest open models
The Artificial Analysis Intelligence Index figures are an independent third-party measurement. Other benchmark figures — DeepSWE improvements, task pass rates, training costs — are Xiaomi's own and have not been submitted to a public leaderboard or independently reproduced. Training cost figures are unaudited. Check the official Xiaomi open-platform documentation for current pricing and specifications before making purchasing decisions.
评论
发表评论