Xiaomi's MiMo-V2.6-Pro: The Best Free Open-Weight AI Model in September 2026

On 22 September 2026, Xiaomi — yes, the phone company — released the strongest open-weight AI model on record. MiMo-V2.6-Pro scores 46 on Artificial Analysis' Intelligence Index, the top open-weights result on that leaderboard, ahead of Z AI's GLM-5.3 (45) and Moonshot's Kimi K3 (44), and level with xAI's proprietary Grok 4.7. Its weights are published on Hugging Face under an MIT licence, which means anyone can download, fine-tune, and run it commercially for free.

That combination — frontier-adjacent scores with zero licence cost — is why this release matters. Here is what MiMo-V2.6-Pro is, what the benchmarks actually show, and where it fits for builders who do not want to pay closed-model API bills.

What was released

The MiMo-V2.6 series arrived as three variants, announced on Xiaomi's MiMo blog on 22 September 2026 (Artificial Analysis lists the release date as 21 September):

• MiMo-V2.6-Pro — the flagship. A sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion activated per token, a 1-million-token context window, and native omnimodal input (text, image, video, and audio) with text output.
• MiMo-V2.6-Flash — the smaller sibling at roughly 310 billion total parameters with 15 billion active per token. Same 1M context window and native multimodal input, targeting high-volume production workloads.
• MiMo-V2.6-Pro-UltraSpeed — a faster variant that Xiaomi says delivers up to 20x faster output at the same quality for latency-sensitive work.

Xiaomi open-sourced the weights for both Pro and Flash under the MIT licence and says it is also releasing the full technical report, the training environments, and the RL code so researchers can reproduce and verify the results.

The benchmark picture: 46 on the Artificial Analysis Intelligence Index

Artificial Analysis' Intelligence Index v4.3.2 is a composite of 10 evaluations, including AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, and Humanity's Last Exam. It is one of the most widely watched independent aggregates because it resists single-benchmark gaming. Here is where MiMo-V2.6-Pro stands, per Artificial Analysis:

Model	Intelligence Index	Licence
MiMo-V2.6-Pro	46	MIT (open weights)
Grok 4.7 (xhigh)	46	Proprietary
GLM-5.3 (max)	45	Open weights
Kimi K3 (max)	44	Open weights
Grok 4.6	44	Proprietary
Gemini 3.8 Flash	41	Proprietary
DeepSeek V4.1 Flash	39	Open weights

Artificial Analysis describes MiMo-V2.6-Pro and GLM-5.3 as the highest-intelligence open-source models. Xiaomi cites a slightly higher 46.32 on the index and calls the model "the strongest open-source model to date".

Xiaomi's own benchmark tables add more texture. On DeepSWE v1.1, a held-out long-horizon software-engineering benchmark, MiMo-V2.6-Pro scores 71.9, up from 19.0 for MiMo-V2.5-Pro — an enormous generational jump, though it still trails DeepSeek V4.1 Flash (74.2), Claude Opus 5 (74.0), and GPT 6 Astra (74.0). Other reported results: 1,673 Elo on GDPVal 2.1 (behind Claude Fable 5.1's 1,735 and Claude Opus 5's 1,708), 76.9 on Toolathlon-verified, 53.1 on Automation Bench v1.0.6, 34.9 on Terminal-Bench 4.0, and 94.0 on CyberGym.

The price-performance story is what really stands out. Artificial Analysis puts MiMo-V2.6-Pro at roughly $0.13 per Intelligence Index task. The closed frontier clusters at $3 to $8 per task for only modestly higher scores — the models at the top of the index (Astra, Opus 5, Fable 5.1) sit in the low-to-mid 50s on raw intelligence.

Under the hood: how Xiaomi got here

According to the model card published by the Xiaomi MiMo team, MiMo-V2.6-Pro-RL uses a 70-layer backbone with 60 sliding-window attention layers interleaved with 10 global-attention layers, 384 routed experts of which eight are active per token. A 681M-parameter MiMo ViT handles vision, a 308M-parameter AudioTokenizer plus a 127M-parameter audio patch encoder handles audio, and a five-layer multi-token-prediction speculative decoder predicts seven subsequent tokens per forward pass. Both models support up to 128,000 output tokens.

The training story is unusual: Xiaomi says it streamed the production reinforcement-learning run live. In under six days, MiMo-V2.6-Flash and MiMo-V2.6-Pro each completed 30 RL steps over roughly 750,000 trajectories, at reported costs of about $0.85 million and $2.62 million respectively. Average pass rate on training tasks rose 25% (Flash) and 12% (Pro) in relative terms, and on the held-out DeepSWE v1.1 the scores climbed from 48.8 to 65.68 for Flash and 58.4 to 72.57 for Pro. The MiMo team is headed by Fuli Luo, a former DeepSeek researcher.

Pricing: the real disruption

API pricing is unchanged from the V2.5 series, so the intelligence gain came free of charge for anyone already calling those endpoints. Per million tokens on Xiaomi's own API platform:

Model	Cache-hit input	Cache-miss input	Output
MiMo-V2.6-Flash	$0.0028	$0.14	$0.28
MiMo-V2.6-Pro	$0.0036	$0.435	$0.87
MiMo-V2.6-Pro-UltraSpeed	$0.036	$4.35	$8.70

Flash at $0.14 in / $0.28 out is remarkably cheap for a 1M-context omnimodal model — the second cheapest major frontier model API worldwide. Artificial Analysis measured about 130 output tokens per second and a 2.17-second time to first token through Xiaomi's API. Both models are available in AI Studio, MiMo Code, MiMo Desktop, the MiMo API Platform, and OpenRouter.

What it's best for

Xiaomi's demonstrations skew toward agentic and creative work: multi-agent game development, Blender 3D modelling, closed-loop control of a simulated Franka Panda robotic arm from multi-view camera feeds, frontend and presentation design, end-to-end video creation, and music composition — including an orchestral piece called Night Road that the model scored and converted to MIDI on its own. In two research cases, MiMo-V2.6-Pro designed metal-organic framework candidates for adsorbing PFAS compounds with Xiaomi's materials experts, and produced a 6,000+ line Lean 4 formalization of Li and Yorke's "Period Three Implies Chaos" theorem, verified by Lean's kernel with no unfinished proof placeholders.

For most builders, the practical answer is simple: Flash is the default for high-volume work — cheap, fast, 1M context, open weights — and Pro is what you escalate to when a task needs the strongest open model available.

Honest limitations

No release is perfect, and this one deserves caveats:

• The closed frontier still leads on raw scores. Astra, Opus 5, and Fable 5.1 sit in the low-to-mid 50s on the index. If your task needs the absolute best reasoning available, the open crown does not close that gap.
• No published SWE-bench number for V2.6 yet. Xiaomi has not published a V2.6 SWE-bench result, so the agentic coding picture is incomplete — DeepSWE 71.9 is good, but trails the proprietary leaders.
• Omnimodal input has a caveat. Xiaomi calls the models "natively omnimodal," but its coding-tool configuration examples list only text and image input. Before you build an audio or video pipeline, check that your API route accepts those inputs.
• "Open weights" is not "free to run." The weights are free, but a 1.02T-parameter MoE still needs serious GPU memory to self-host. Most people will use it through an API, where the pricing is cheap but not free.
• Vendor benchmarks need independent confirmation. Some numbers above (DeepSWE gains, demo results) come from Xiaomi's own reporting. Independent replication is pending, which is exactly why Xiaomi open-sourcing the training environments matters.

How to try it without paying anything

If you just want to play with a strong open-weight model in a chat interface, you have options. The weights are on Hugging Face (MIT licence), the models are on OpenRouter, and Xiaomi offers AI Studio and MiMo Desktop.

And if you simply want free AI chat in the browser with no signup — the same habit this whole free-AI trend runs on — Toolxz AI Chat gives you free unlimited frontier AI chat without logging in, alongside 45+ other free browser-based tools. It is a handy fallback whenever an API is rate-limited or a demo page asks you to create an account first.

If you are comparing open models more broadly, my earlier post Free Bilingual AI Chat (English + Chinese): GLM 5.2, No Signup covers another strong open model family — GLM-5.3 is now MiMo-V2.6-Pro's closest rival on the open-weights leaderboard.

FAQ

Is MiMo-V2.6-Pro really free?
The weights are free under the MIT licence — download, fine-tune, deploy commercially, no per-token fee. Running them yourself needs powerful hardware, though; the hosted API starts at $0.14/$0.28 per million tokens for Flash.

Is it better than DeepSeek?
On the Artificial Analysis Intelligence Index, yes: MiMo-V2.6-Pro scores 46 versus DeepSeek V4.1 Flash at 39. It also tops Kimi K3 (44) and GLM-5.3 (45).

Can it beat GPT or Claude?
Not on raw benchmark scores — the top proprietary models sit in the low-to-mid 50s. But it matches Grok 4.7 (46) at a fraction of the API cost, and you can run it yourself.

What does "omnimodal" mean?
The model accepts text, image, video, and audio as input and outputs text. Practically, check your API route — some endpoints currently accept text and image only.

Who is behind it?
Xiaomi's MiMo team, headed by former DeepSeek researcher Fuli Luo. It is the team's third major open-weight release in roughly a year.

The bottom line

MiMo-V2.6-Pro is the strongest downloadable model you can get right now, at an API price that undercuts almost everything in its performance tier. The 46 on the Intelligence Index ties Grok 4.7 — a proprietary model from xAI — while remaining MIT-licensed and free to self-host. The $2.62 million RL run bought roughly a 4-point index jump for the V2.5 series' successor, and Xiaomi kept the price flat.

The open-weights race is now genuinely competitive with the closed frontier on value. Whether it stays that way depends on independent replication — and Xiaomi releasing the training environments is the right move to make that possible.

Sources: Unite.AI, VentureBeat, i-scoop.eu, mixed-news.com, eesel.ai. Benchmark data: Artificial Analysis Intelligence Index v4.3.2.

Founder of Toolxz (toolxz.com) — 45+ free browser-based tools. I write about practical AI tooling and developer workflows.