MiMo V2.6 Flash and Pro Are Now on Multos — Frontier-Grade Agents at Flash-Tier Pricing
Xiaomi released MiMo V2.6 Pro and Flash on September 21, 2026. Pro scored 46 on the Artificial Analysis Intelligence Index — the highest any open-weight model has ever scored, at $0.13 per task. Flash costs $0.14/$0.28 per million tokens and trails Pro by only 4 points on DeepSWE. Both are live on Multos Lite via DeepInfra.
Xiaomi released MiMo V2.6 Pro and Flash on September 21, 2026, and both are available on Multos Lite via DeepInfra starting today. These are the most capable open-weight models available right now — and both are priced in the flash tier, not the premium tier. Pro scored 46 on Artificial Analysis's Intelligence Index, the highest any open-weight model has ever scored. More importantly, Artificial Analysis puts its cost per task at $0.13 — less than GPT-5.6 Sol at $1.99, less than Kimi K3 at $2.00, less than Claude Opus 5 at $5.86. Flash trails Pro by only 4 points on DeepSWE v1.1 and costs roughly a third as much per token: $0.14 input, $0.28 output. Both models support a 1-million-token context window and take text, images, video, and audio as input.
What MiMo V2.6 Actually Is
MiMo V2.6 is a pair of Mixture-of-Experts models. Pro has 1.02 trillion total parameters with 42 billion active per token. Flash has 309 billion total with 15 billion active. Both were pretrained at scale — Flash on 48 trillion tokens, Pro on 30 trillion — then extended to a 1-million-token context during mid-training, then put through a large-scale reinforcement learning run. That RL run is the unusual part. Xiaomi ran 30 update steps per model, generating 25,088 agent trajectories per step — roughly 750,000 long-horizon trajectories per model total. Each trajectory averaged 110,000 to 150,000 tokens: read a codebase, search for files, call tools, rewrite code, run tests, observe failures, revise. Not a question-answer pair. A full work session. The RL cost was approximately $2.62 million for Pro and $850,000 for Flash. More than half of that went to generating and grading those trajectories — not to the parameter updates themselves. Xiaomi introduced a groupwise reward system that compares multiple trajectories from the same task and steers the model toward shorter, more efficient solutions rather than just successful ones. That design philosophy shows up in the benchmarks: MiMo V2.6 is efficient in a way that pure benchmark scores don't fully capture. Weights are MIT-licensed. You can download, fine-tune, and deploy them commercially. Xiaomi also released 7,000+ RL task environments, an end-to-end RL framework, and a distilled 9B model — MiMo-V2.6-Distill-Qwen-9B — for researchers who want to replicate the training approach at smaller scale.
The Benchmarks
From Xiaomi's model card, comparing Pro and Flash against Claude Opus 5 and GPT-5.6 Sol: DeepSWE v1.1 (long-horizon software engineering): Pro: 71.9% | Flash: 67.9% | Claude Opus 5: 74.0% | GPT-5.6 Sol: 73.0% Pro and Flash are close behind the closed frontier models on the hardest long-horizon coding benchmark available. The gap is 2-6 points depending on which closed model you compare to. AutomationBench v1.0.6 (business workflow automation): Pro: 53.1% | Flash: 52.3% | this benchmark: Pro and Flash beat Claude Opus 5 and GPT-5.6 Sol Terminal Bench 2.1: Pro: 89.9% | Flash: 87.6% MiMo Visual Coding: Pro: 72.3% | Flash: 71.5% — native visual input for UI and frontend work The Artificial Analysis Intelligence Index score of 46 for Pro is the headline, and it's real — verified by an independent third party. But the more honest framing is that MiMo V2.6 Pro is the top open-weight model, not the overall AI leader. Claude Opus 5 max scored 51 on the same index. GPT-5.6 Sol max scored 47. MiMo V2.6 Pro at 46 sits right behind them — at $0.13 per task versus $5.86 and $1.99 respectively. Flash at $0.14/$0.28 per million tokens and similar performance on most evaluations is the more interesting release for production use. The 4-point gap on DeepSWE, the 0.8-point gap on AutomationBench, and the 2.3-point gap on Terminal Bench 2.1 are all narrow. The price gap between Flash and Pro is not: Flash is about 32% of Pro's cost per token. For agent workloads where the model makes dozens of calls per session, that ratio compounds.
Building Apps with MiMo V2.6 Flash on Multos
Flash's combination of a 1-million-token context window, native multimodal input, strong agentic benchmarks, and flash-tier pricing opens up some specific workflows on Multos that weren't previously available at Lite. Full-repository analysis. Paste an entire codebase — even a large one — into context without chunking. At 1M tokens and 15B active parameters, Flash can hold the entire project in one context and reason about cross-file dependencies. For app builders: ask it to audit your data model, find inconsistencies between schema and code, and propose migrations — all in a single call with full context. Visual UI work. MiMo V2.6 takes images and video natively, not as a bolt-on. Paste a screenshot of your app's interface and ask what's broken. Share a screen recording and ask where users might get confused. This works at the Flash tier on Multos Lite — you don't need a Pro or Elite model to get native visual understanding. High-volume agents. Flash at $0.14/$0.28 per million tokens makes multi-step agents viable at scale on Lite. If you're building a tool that runs many agent turns per user session — a coding assistant, a data pipeline, a document processor — the per-call economics at Flash pricing mean your pool goes much further per meaningful unit of work than with the standard mid-tier models. Game development: Flash's 71.5% on MiMo Visual Coding means it understands game asset screenshots, UI layouts, and rendered game states. Use it for the implementation and debugging loop — feed it screenshots of the running game alongside the relevant code and it can reason about both simultaneously.
Pro vs Flash — When to Use Each
For most Multos workloads, Flash is the better default. The performance gap is narrow on the evaluations that reflect real building tasks — a few points on coding benchmarks, sub-1-point on business automation. The cost difference is substantial. Use Flash for: routine app development, debugging sessions, UI work with screenshots, any pipeline that makes many model calls, context-heavy tasks that benefit from the 1M window, and anything where you want to try multiple approaches before committing to one. Use Pro for: the most demanding long-horizon coding tasks where DeepSWE-level performance matters, tasks where you need maximum capability on agentic terminal benchmarks (34.9% vs 28.8% on Terminal Bench 4.0), and security-oriented work where the ExploitBench gap (47.9% vs 25.3% for Flash) is relevant. A practical escalation pattern: default to Flash, route to Pro when Flash's first attempt doesn't meet quality, and use Pro directly only for the class of tasks where you know up front that Flash won't be sufficient. This mirrors how high-volume production systems handle model selection and keeps costs in check without sacrificing capability on the hard cases.
DeepInfra — Why It Matters
Both MiMo V2.6 models on Multos are served via DeepInfra. DeepInfra listed them today — September 22 — literally within hours of Xiaomi's release. They're in DeepInfra's 'Latest Models' section with live pricing and a public API. DeepInfra uses fp8 precision for serving and zero data retention. For Multos users: the connection is managed — you don't configure anything, no API key, no endpoint setup. You pick MiMo V2.6 Flash or Pro from the dropdown and Multos handles the routing. Fallback is built in: if DeepInfra is under load, Multos routes to the novita MiMo V2.5 Pro as a fallback for Flash, and keeps the session alive. As other providers add V2.6 in the coming days — Novita, Fireworks, Baseten, and Together have all historically listed MiMo within days of a new release — the fallback pool will grow.
Ready to Get Started?
MiMo V2.6 Flash and Pro are live on Multos Lite via DeepInfra. The top open-weight model in the world, at flash-tier pricing. Select either from the model dropdown under DeepInfra — no setup, no API key, available today on every paid Multos plan.
Start Building FreeRelated Articles
Claude Opus 5.5 Is Live on Multos — Fable-Class Work at 40% Lower Cost
Anthropic released Claude Opus 5.5 today and it's immediately available on Multos Elite. It matches Claude Fable 5.1 on most tasks, costs 40% less to run than Opus 5, and is 30% faster. On Terminal-Bench 4.0 it scores 66.4% — higher than GPT-6 Astra, higher than Opus 5, higher than Fable 5.1. Here's what that actually means for building apps and games on Multos.
Read more →GPT-6 Sol and GPT-6 Luna Are Live — OpenAI Halves the Price of Its Best Models
OpenAI released GPT-6 Sol and GPT-6 Luna today, cutting prices by 50% versus GPT-5.6. Sol drops from $4/$20 to $2/$10 per million tokens. Luna drops from $0.20/$1.20 to $0.10/$0.50. Intelligence Index scores hold steady, hallucinations drop significantly, and both models are live on Multos now — Sol on Pro, Luna on Lite.
Read more →GLM-5.3 Flash: The Model That Was Hiding in Plain Sight
For six days, one of the most capable models on OpenCode and OpenRouter had no name. It was just called Ox Alpha. Then Z.ai revealed it: GLM-5.3 Flash, a 320B MoE model with a 1M token context, native multimodal input, MIT-licensed weights — and a price tag that makes you read it twice.
Read more →