Xiaomi's new open AI model tops the open-weight rankings

On September 21, 2026, Xiaomi released and open-sourced the MiMo-V2.6 series — three checkpoint families headed by two natively omnimodal sparse-MoE models, published as ungated, MIT-licensed Hugging Face repositories:

• MiMo-V2.6-Pro-RL — flagship: 1.02T total / 42B active parameters, 1M-token context, text/image/video/audio input, text output. (FACT, HF card + AA agree)
• MiMo-V2.6-Flash-RL — efficiency tier: 309B total / 15B active, same 1M-token context and modalities. (FACT)
• MiMo-V2.6-Distill-Qwen-9B — a 9B agentic-RL starting point distilled on MiMo-generated data. (FACT, HF; company describes it as a research entry point)
• Plus 7,000+ high-quality RL task environments (software engineering, vulnerability reproduction, knowledge-intensive work, web design & development), an end-to-end RL training framework (built on verl, uni-agent, mini-swe-agent) and composable mini-harnesses — i.e., Xiaomi opened the training machinery, not just the weights. (COMPANY CLAIM as to coverage/quality; release contents CONFIRMED from official docs)
• The official announcement (dated Sep 22) is titled "MiMo-V2.6: Scaling Up Reinforcement Learning for Self-Improvement," with the blog page meta: "Introducing the MiMo-V2.6 series: frontier intelligence, all the modalities, built in public."

Independent measurement (INDEPENDENTLY VERIFIED): Artificial Analysis scores MiMo-V2.6-Pro at 46 on the Intelligence Index v4.3.2 — the highest score of any open-weights model on its leaderboard, #1 of 114 in its class (large open-weights reasoning models), tied overall with the proprietary Grok 4.7 (xhigh). AA lists the release date as September 21, 2026; API pricing $0.435/M input, $0.87/M output with an unusually deep 99% cache discount, $0.13 per Intelligence Index task (vs ~$3.74 for Grok 4.7), 54.2 t/s, TTFT 2.67 s, 1M context, and a single API provider (Xiaomi's own).

Company-published performance claims (COMPANY CLAIM until replicated): Pro reaches 71.9 on DeepSWE v1.1 (vs 67.9 Flash, 19.0 MiMo-V2.5-Pro, 74.0 Claude Opus 5 / GPT-5.6 Sol, 70.0 Claude Fable 5); on most agent benchmarks Xiaomi says Pro is on par with Claude Opus 5 and GPT-5.6 Sol; AA index "46.32" (per Unite.AI's quote of the announcement); the series "surpasses Kimi K3 and Qwen3.8 Max"; multimodal capabilities across 3D game generation, Blender modeling, Computer Use, embodied control of a Franka Panda arm, materials research (MOF design for PFAS adsorption), and a Lean 4 kernel-verified formalization of Li & Yorke's "Period Three Implies Chaos" (>6,000 lines).

Production RL transparency — the unusual part: Xiaomi livestreams its production RL runs. In under six days the Flash and Pro checkpoints each completed 30 RL steps over ~750,000 trajectories total, at reported costs of ~$0.85M (Flash) and ~$2.62M (Pro); average pass rate on training tasks rose 25% (Flash) and 12% (Pro) and DeepSWE v1.1 (a held-out long-horizon benchmark) improved from 48.8 → 65.7 (Flash) and 58.4 → 72.6 (Pro). (COMPANY CLAIM — figures published in the official announcement; costs not auditable from outside.)

Availability (FACT): API model names mimo-v2.6-pro, mimo-v2.6-flash, mimo-v2.6-pro-ultraspeed (all-lowercase); API prices unchanged from V2.5; an UltraSpeed mode (up to 20× inference speed) at 10× the token price; Xiaomi MiMo Desktop client official release with membership; also in AI Studio, MiMo Code, and OpenRouter, plus ModelScope mirrors.

• The strongest open-weights release ever measured on AA's Intelligence Index landed — with the training stack attached. MiMo-V2.6-Pro's 46 beats the previous open-weights frontrunners (GLM-5.3 max 45, Kimi K3 max 44) on an independent leaderboard, and Xiaomi open-sourced not only MIT weights but 7,000+ RL environments, the RL framework, and mini-harnesses. (FACT on the score; scope of the open-sourced machinery per official docs, CONFIRMED)
• The open/closed intelligence gap on AA's composite index closed to one point for the top open model vs the top proprietary field ordering: open-weights 46 vs Claude Fable 5.1 / GPT-6 Astra at 53 — and MiMo ties the proprietary Grok 4.7 at 46. The "open weights are one generation behind" assumption no longer holds on this composite. (INTERPRETATION grounded in AA data)
• Licenses shifted: the earlier MiMo-V2-Flash repo was Apache-2.0; the V2.6 cards are MIT (FACT: card YAML license: mit), removing even attribution friction for commercial derivatives.
• RL training economics went public and quasi-reproducible: ~$2.62M for a 6-day, 30-step production RL run on the flagship, with the environment/framework code released so others can reproduce the loop — a step-change in transparency for frontier-scale RL. (COMPANY CLAIM on the costs; reproducible via the released framework — EARLY RESEARCH)
• A new price-performance reference point for open weights: AA measures $0.13 per Intelligence Index task for Pro at $0.435/$0.87 per M with a 99% cache discount — far below the ~$3.7/task of Grok 4.7 and ~$3.3/task of GPT-6 Astra Max, i.e., per-intelligence-point cost collapsed by ~1–2 orders of magnitude. (INDEPENDENTLY VERIFIED on the AA numbers; the cross-vendor comparison is arithmetic on AA data)
• Hardware-adjacent pressure: a 1T-parameter MoE with 1M-token training at 3.5–3.7B tokens/step (company claim) raises the baseline for what "open-source AI" requires to run — and makes Flash (15B active) the practical on-prem entry point.

Before (as of mid-September 2026):

• Open-weights leadership on AA's Intelligence Index: GLM-5.3 (max) at 45, Kimi K3 (max) at 44 — no open model had crossed 46; MiMo-V2.5-Pro sat at 19.0 on DeepSWE v1.1 with no competitive agentic profile (FACT: Xiaomi's own before/after table).
• Open-weights releases typically shipped weights + a card + maybe a small eval harness; training environments and RL code stayed proprietary. (INTERPRETATION, widely observable)
• Xiaomi MiMo's earlier open releases (MiMo-V2-Flash) were Apache-2.0, and V2.5 was open-sourced with an "Orbit 100 trillion token plan" (from the docs news index).
• Closed frontier: Claude Fable 5.1 / GPT-6 Astra at 53 holding a ≥7-point edge over any open model; pricing war in full swing (Grok 4.7 same week at $2/$6; GPT-6 Sol/Luna the next day at half price).

Change (Sep 21–22, 2026):

• MiMo-V2.6-Pro-RL (1.02T/42B, MIT, ungated) hits 46 on AA — #1 open-weights score, tied overall with proprietary Grok 4.7; Flash (309B/15B) ships the same 1M context and modalities at 10× lower token price than Pro.
• Xiaomi releases weights plus 7,000+ RL environments, an end-to-end RL framework, mini-harnesses, a technical report, and a Distill-9B starting point, with a livestreamed 6-day production RL run documented end-to-end.
• API pricing unchanged from V2.5; UltraSpeed (≤20×) launched; Desktop client leaves beta; Distill-9B demonstrates 61.1 → 66.2 on SWE-bench Verified after RL (company claim).

After:

• For the first time, the top open-weights model and a top proprietary model sit at the identical composite index score (46/46) out of 673 tracked models, while the overall lead (53) remains closed — so the open-vs-closed narrative splits into "parity at the top of open" vs "overall lead still closed."
• Every lab's next release must now answer "open at 46 with the RL stack attached" — either on score, on price-per-task ($0.13), on license (MIT), or on transparency (livestreamed RL runs). (PREDICTION)
• Developer default deployment choice for cost-sensitive agentic work shifts toward open weights: MiMo-V2.6-Flash (15B active, $0.14/$0.28) becomes a credible self-hosted default, with Pro as the moat-level option. (INTERPRETATION)

• Architecture (FACT, HF cards): both checkpoints are sparse MoE on a hybrid sliding-window/global-attention backbone with "You Only RL Once" mixed training. Pro: 70 layers (60 SWA + 10 GA), hidden 6144, 384 routed experts with 8 active per token, SWA window 128, 1M max context. Flash: 48 layers (39 SWA + 9 GA), hidden 4096, 256 experts / 8 active. First block uses global attention with a dense FFN; remaining blocks interleave local SWA and GA with sparse MoE FFNs and no shared experts.
• Omnimodal I/O (FACT): a 681M-param MiMo ViT (28 layers, 24 SWA + 4 full; patch 2×16×16) for vision; audio via a 308M-param AudioTokenizer (20 RVQ codebooks) plus a 127M-param audio patch encoder (25 Hz → 6.25 Hz); text/image/video/audio input, text output; 1M-token context (about 1,500 A4 pages per AA).
• Speculative decoding (FACT): a 5-layer SWA multi-token-prediction drafter predicts 7 subsequent tokens per forward pass for parallel verification (EAGLE-style recipes in the cards).
• RL recipe (COMPANY CLAIM as to efficacy; mechanics from cards/docs): fully asynchronous GRPO on huge batches — 1,568 prompts × 16 rollouts per step, billions of tokens per update, 1M-context training with 3.5–3.7B tokens per step. Grading is itself scaled: Groupwise Reward Synthesis (GRS) builds task-specific rubrics offline by contrasting rollouts; Groupwise Advantage Redistribution (GAR) ranks passing trajectories online — together forming a self-improvement loop that steers toward shorter paths and fewer tokens.
• Stability/anti-hacking measures (COMPANY CLAIM): the MoE router is frozen during RL to suppress expert load drift; a reward-hacking defense spans reward design, adversarial evaluation, anomaly detection, and verifier cross-checks.
• Multi-harness design (COMPANY CLAIM): tasks and harnesses (code, general agents, visual, cyber) are mixed in the same batch so capabilities reinforce each other; lightweight composable "mini-harnesses" decouple system prompts, tools, and context management, and Xiaomi claims generalization to harnesses never seen in training.
• Distillation (COMPANY CLAIM): MOPD2 (Multi-Prefix Multi-Teacher On-Policy Distillation) follows the mixed RL run, reusing teacher/SFT prefixes so decision points train without regenerating preceding turns — extending RL gains to hard-to-verify tasks; the Distill-Qwen-9B artifact applies this to a small startable checkpoint.
• Serving (FACT): SGLang and vLLM recipes published on the model cards (Docker images, tensor-parallel/EP layouts); also served via Xiaomi's API (OpenAI-compatible), OpenRouter, AI Studio, MiMo Code, and Desktop; UltraSpeed mode at up to 20× token speed at 10× price.

1. The open-weights ceiling moved — and the gap to closed models narrowed dramatically on a respected composite. AA 46 (#1 open, tied with proprietary Grok 4.7; closed leaders at 53) is the strongest independently measured open-weights result to date. The "open weights are always a generation behind" claim is no longer true at the top of the open field. (FACT on scores; the claim about what it means is INTERPRETATION)
2. Xiaomi open-sourced the reproduction kit, not just the model. 7,000+ RL environments, the verl-based end-to-end framework, mini-harnesses, a technical report, and a 9B starting point mean other labs and enterprises can rebuild the training loop — this is a public-good move that could accelerate open agentic-RL research globally. (FACT on the release scope; direction of effect is PREDICTION)
3. Per-task cost collapse for open agentic AI: $0.13 per Intelligence Index task (vs $3.74 Grok 4.7, ~$3.3 GPT-6 Astra Max) resets the economics of open-weights agentic workloads — with a 99% cache discount that rewards long-context/RAG patterns. (INDEPENDENTLY VERIFIED numbers; the reset is INTERPRETATION)
4. The pricing war now has an open-weights supply side: same week as Grok 4.7 ($2/$6) and GPT-6 Sol/Luna (half price), Xiaomi shipped MIT weights at $0.435/$0.87 (Pro) and $0.14/$0.28 (Flash) with unlimited commercial use — "frontier for less" is now a three-company reality plus a free-for-use open layer. (FACT)
5. Transparency as a competitive weapon: livestreamed production RL, published run costs, and full training-code release is a genuinely different posture from closed labs — it converts Xiaomi's AI research into a trust signal and a recruiting/ecosystem asset. (INTERPRETATION)

• Self-hosting a top-of-open-weights reasoning model at 1M context with native text/image/video/audio input — no API dependency, no per-token bill, MIT terms — for organizations that can provision the hardware (Pro) or that pick Flash (15B active) for feasible on-prem inference.
• Reproducing and extending a frontier-scale agentic RL loop: the released environments (SW engineering, vulnerability reproduction, knowledge-intensive work, web design/dev) + framework + mini-harnesses give research teams a starting point they previously had to assemble from scratch (months of work). (COMPANY CLAIM on quality/completeness; contents CONFIRMED)
• Running long-horizon agent workloads (1M context; per AA "~1,500 A4 pages") on open weights with a 99% cache discount — RAG over large codebases, multi-session agent runs, tool traces.
• Building commercial products on MIT weights with zero license friction (no attribution clause, unlimited redistribution/modification), including fine-tunes, distillation, and serving businesses.
• Small-team agentic-RL research via Distill-Qwen-9B: a 9B checkpoint documented to gain +5.1 SWE-bench Verified and +15.7 MiMo Cyber Bench from RL over its SFT baseline (company claims) — a startable, GPU-friendly training target.
• Benchmarking openness: AA's Openness Index components and license data now include a model that is both top-intelligence and maximally permissive — a reference point for procurement "openness" requirements.

1. Will independent re-runs confirm DeepSWE 71.9 (Pro), CyberGym 94.0, and the agent-table numbers — and where do they land on AA's Coding Agent Index once measured (not yet on the fetched pages)?
2. Do the claimed RL-economics (6 days, 30 steps, ~750K trajectories, $2.62M for Pro) hold up to reproduction with the released framework — and what cluster was required?
3. Why does the HF API return license: None while cards/README/AA all say MIT — metadata migration bug or intentional flag? Will automated license audits be misled?
4. Is UltraSpeed the documented 129.7 t/s readout (early press) vs the current 54.2 t/s default — and does the 20× claim hold at 1M context?
5. When will third-party providers (Fireworks/Together/Baseten-class) host Pro/Flash, and at what prices relative to Xiaomi's $0.435/$0.87?
6. Will Xiaomi release the base (non-RL) V2.6 checkpoints or a V2.6-Pro "non-reasoning" variant for latency-sensitive serving?
7. What do independent safety/red-team evaluations of the released cyber environments conclude — and does the open cyber kit attract researcher scrutiny or misuse first?
8. Does the Distill-Qwen-9B recipe (SWE-bench 61.1→66.2, MiMo Cyber Bench 31.3→47.0) reproduce on other 9B bases, and does it generalize outside MiMo's environment suite?
9. What is the roadmap cadence now — MiMo-V2.7? Integration with MiMo Code/Claw/Orbit ecosystem — and does the API pricing really stay flat?

• Days: AA leaderboard refresh (open-source + overall) as more of the 673-model index runs land for Pro/Flash; possible third-party provider announcements; community re-runs of DeepSWE/agent tables begin (early HF discussion threads exist — 8 discussions on Pro, 7 on Flash).
• Weeks: first independent replication attempts of the RL recipe (environments + framework are fresh downloads); Xiaomi Orbit/Claw/Desktop ecosystem integrations; UltraSpeed beta ends (invitation-only program runs "one more week" per the docs) — watch pricing stability after the beta window; possible base/non-RL variant release.
• Months: open-weights leaderboard consolidation (MiMo 46 vs GLM-5.3 45 vs Kimi K3 44) as each lab replies; closed-lab budget-tier responses to the $0.13/task cost line; Q4 procurement cycles incorporating MIT-open-weights options; and at some point MiMo-V2.7-era RL-scale transparency becoming the norm for major open releases (PREDICTION).

Xiaomi's MiMo-V2.6 release is the strongest open-weights story of this window for two independent reasons: score — 46 on Artificial Analysis' Intelligence Index, the #1 open-weights result ever measured there, verified by the benchmark owner itself, tied with the proprietary Grok 4.7 and only 7 points behind the closed leaders; and reproduction kit — MIT weights plus 7,000+ RL environments, the full training framework, mini-harnesses, a technical report, and a livestreamed $2.62M RL run that anyone can now attempt to rebuild. The honestly-labiled framing: Xiaomi's own flagship benchmark bullets (DeepSWE, agent tables, the "surpasses Kimi K3 and Qwen3.8 Max" line, the 1/20–1/60 price ratio) are COMPANY CLAIMS awaiting independent replication, and 46 is the open-weights top, not the overall top (Fable 5.1 / GPT-6 Astra lead at 53). The through-line for the weekly: the frontier-for-less pricing war now has an open-weights supply side at $0.13 per index task with a 99% cache discount, and "open-sourcing the training machinery" may matter more long-term than any single benchmark row.