--- title: MiMo-V2.6-Pro tops open weights models on Artificial... description: MiMo-V2.6-Pro has claimed the top spot among open weights models on the Artificial Analysis Intelligence Index with a score of 46. It keeps the same... canonical: https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: MiMo-V2.6-Pro tops open weights models on Artificial Analysis Intelligence Index | daily.dev og:description: MiMo-V2.6-Pro has claimed the top spot among open weights models on the Artificial Analysis Intelligence Index with a score of 46. It keeps the same... og:url: https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee og:image: https://api.daily.dev/og/posts/oYEVo1PeE.png og:image:alt: MiMo-V2.6-Pro tops open weights models on Artificial Analysis Intelligence Index og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # MiMo-V2.6-Pro tops open weights models on Artificial Analysis Intelligence Index **[Collections](https://daily.dev/sources/collections)** · 3 min read · 4 upvotes · 0 comments ## Summary MiMo-V2.6-Pro has claimed the top spot among open weights models on the Artificial Analysis Intelligence Index with a score of 46. It keeps the same architecture as V2.5 but adds a 1M token context window and an MTP drafter, landing just behind GPT-4.5 Sol while costing $0.13 per million input tokens. ## Content Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash last week, and the reaction from the ML community has been unusually enthusiastic — not just for the benchmark numbers, but for how much Xiaomi actually published alongside the weights. ## What was released Both models are MIT-licensed and multimodal (text, image, video, audio in a single checkpoint), with a 1M-token context window and native FP8 weights. The architecture specs: - **Pro**: 1.02T total parameters, 42B active - **Flash**: 309B total parameters, 15B active Both use the same architecture as V2.5 — hybrid sliding-window plus global attention, with DFlash speculative decoding that drafts 7 tokens per step. vLLM shipped day-0 support for both sizes. Beyond the weights, Xiaomi also released: - Over 7,000 graded RL task environments covering code, math, and agent tasks - The full end-to-end RL training framework - A detailed technical report with hyperparameters and cost data - A smaller distilled model, MiMo-V2.6-Distill-Qwen-9B Training cost for both variants combined came to roughly $3.5 million, completed in under six days. Xiaomi livestreamed the RL runs publicly for five days. ## Benchmark position Artificial Analysis ranks MiMo-V2.6-Pro first among 114 tracked open-weight models on their Intelligence Index, scoring 46. At $0.13 per task, it's also the cheapest model they track — well below Anthropic's Claude Opus 5.5 even after recent price cuts. That said, Pro still trails closed frontier models on several benchmarks, particularly Terminal Bench 4.0. Flash, priced at roughly a third of Pro's cost, scored slightly higher than Pro in at least one hands-on coding comparison (72.5% vs 69.38% across eight tasks), making it the more practical production choice for most use cases. ## Why the architecture is interesting MiMo-V2.6 uses Grouped Query Attention (GQA) with a 128-token Sliding Window Attention window — nothing exotic. Researcher Sebastian Raschka pointed out that this is the point: the model sits at number one on open-weight benchmarks despite a straightforward attention design, which suggests most of the gains came from data and post-training improvements rather than architectural novelty. The technical report details three things that stood out: 1. **More agent tasks, across more harnesses.** DeepSWE pass@1 on held-out harnesses improved from roughly 50% to 66%. 2. **Better reward signals.** A simple correctness verifier was replaced with an agentic grader that examines full execution traces. 3. **Large RL batches.** 1,568 prompts × 16 rollouts = 25,088 trajectories per update, with 2.7–3.7 billion training tokens per update. ## The openness angle The RL environments release is what got the most attention from researchers. As training increasingly shifts toward reinforcement learning with verifiable rewards (RLVR), curated task environments are becoming genuinely valuable — and most labs keep them private. Hugging Face's Thomas Wolf and others specifically called out the environment release as the notable part of this drop. This contrasts with recent open-weight releases from Meta, Moonshot, and Alibaba, which published weights but little else about the training process. Peano AI has already run full-parameter RL on the 310B Flash model across 1,000+ TPUs using vLLM for rollouts, with weights syncing across the ICI fabric in under two seconds. ## Known issues A few bugs surfaced in early testing: Pro mishandled elevator state logic and clock-hand positioning in coding tasks, while Flash hit output token limits on a permutation math problem and had its own clock-positioning errors. The MiMo team acknowledged and patched a tool-call repetition bug shortly after launch. One other note: Xiaomi is separately named in an Anthropic report alleging misuse of Claude's output, though that hasn't been addressed publicly by either party in the context of this release. ## Questions this post answers ### What score did MiMo-V2.6-Pro get on the Artificial Analysis Intelligence Index? MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, making it the top-ranked open weights model on that benchmark. It sits just behind GPT-4.5 Sol in overall ranking while being priced at $0.13 per million input tokens, a notable cost-to-performance position for an open weights release. _Developers comparing open weights model pricing and benchmarks can track releases like this one on daily.dev._ ### What changed in MiMo-V2.6-Pro compared to MiMo-V2.5? MiMo-V2.6-Pro keeps the same underlying architecture as V2.5 but adds a 1M token context window and an MTP (multi-token prediction) drafter. These additions helped push it to the top of the open weights ranking on the Artificial Analysis Intelligence Index with a score of 46. _Teams evaluating context window upgrades across model releases can follow updates like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 8 comments across x (as of 2026-09-28). **TL;DR:** Reactions focus on Xiaomi's release of RL training environments and reward data alongside the model weights, which people see as more valuable and rarer than the weights themselves. **Sentiment:** 65% positive · 30% mixed · 5% skeptical **The case for** - Releasing RL task environments with ground truth reward models is seen as a genuine moat-breaking move, rarer than open weights alone. - Some frame the RL dataset as 'the recipe' rather than just a snapshot, making it more reproducible and useful for others. **The pushback** - One commenter argues the real value depends on whether baselines, data licensing, contamination checks, and generalization results are also disclosed, not just dataset scale. **By community** - x (positive): Replies mostly praise the transparency of releasing RL environments and training data alongside the weights. **Open questions** - Will Xiaomi publish baselines, data licensing details, contamination checks, and generalization results on unseen tasks alongside the RL dataset? **Highlights** > @iScienceLuvr Open weights are everywhere now, but open RL environments with ground truth reward models? That's the scarce half. Most labs treat this stuff as their moat. Respect to Xiaomi for actually releasing it. > — [econ\_tech\_vance on x](https://x.com/econ_tech_vance/status/2103737834482086075) > @iScienceLuvr weights are a snapshot, the rl dataset is the recipe. releasing the recipe is the move that actually matters > — [BuilderCulture\_ on x](https://x.com/BuilderCulture_/status/2103741073529213218) > @iScienceLuvr 这类数据集的价值不只在规模，而在能否让 RL 训练从“跑过一次”变成可复现实验。真正决定工程收益的是任务定义、奖励信号和训练后泛化：最好同时公开基线、数据许可、污染检查，以及在未见任务上的提升。 > — [Grok9527 on x](https://x.com/Grok9527/status/2103762859964678344) **Source threads** - [x](https://x.com/iScienceLuvr/status/2103729680960590050) · 0 points · 8 comments ## Similar posts on daily.dev - [Xiaomi releases MIT‑licensed MiMo models for long‑running AI agents](https://daily.dev/posts/xiaomi-releases-mit-licensed-mimo-models-for-long-running-ai-agents-qgbigdop8) · InfoWorld · 2 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"MiMo-V2.6-Pro tops open weights models on Artificial Analysis Intelligence Index","url":"https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee"},"datePublished":"2026-09-21T21:03:03.158Z","dateModified":"2026-09-28T00:15:59.876Z","description":"MiMo-V2.6-Pro has claimed the top spot among open weights models on the Artificial Analysis Intelligence Index with a score of 46. It keeps the same...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"MiMo-V2.6-Pro tops open weights models on Artificial Analysis Intelligence Index"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/mimo-v2-6-pro-tops-open-weights-models-on-artificial-analysis-intelligence-index-oyevo1pee#faq","mainEntity":[{"@type":"Question","name":"What score did MiMo-V2.6-Pro get on the Artificial Analysis Intelligence Index?","acceptedAnswer":{"@type":"Answer","text":"MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, making it the top-ranked open weights model on that benchmark. It sits just behind GPT-4.5 Sol in overall ranking while being priced at $0.13 per million input tokens, a notable cost-to-performance position for an open weights release. Developers comparing open weights model pricing and benchmarks can track releases like this one on daily.dev."}},{"@type":"Question","name":"What changed in MiMo-V2.6-Pro compared to MiMo-V2.5?","acceptedAnswer":{"@type":"Answer","text":"MiMo-V2.6-Pro keeps the same underlying architecture as V2.5 but adds a 1M token context window and an MTP (multi-token prediction) drafter. These additions helped push it to the top of the open weights ranking on the Artificial Analysis Intelligence Index with a score of 46. Teams evaluating context window upgrades across model releases can follow updates like this on daily.dev."}}]} ```