FULL STORY

Xiaomi MiMo-V2.6: Open-Sourced and Topping the Charts

Xiaomi live-streamed MiMo's RL training, then open-sourced the MiMo-V2.6 family under MIT, with the Pro model topping open-source leaderboards and impressing in hands-on tests.

2026-09-21 ~ 2026-09-24 · 4 episodes · 58 posts

Episode 1 · Xiaomi MiMo Finishes RL Training, DeepSWE Score Jumps to 72.57 (2026-09-21, 2 posts)

Xiaomi's MiMo completed reinforcement learning training, boosting its DeepSWE score from 58.41 to 72.57—near the 74 record—while MiMo-V2.6 emerged as the latest open-source model to publicly document its RL training.

Episode 2 · Xiaomi open-sources MiMo-V2.6, topping open-weight leaderboard at 46 with livestreamed RL training (2026-09-22, 47 posts)

On September 22, Xiaomi open-sourced the MiMo-V2.6 fully multimodal model series on Hugging Face under the MIT license. The Pro model scored 46 on the Artificial Analysis Intelligence Index, topping the open-weight leaderboard for the first time—a large jump from MiMo-V2.5-Pro's 26 and only about 1 point behind GPT-5.6 Sol. The team also released the tech report, RL environments and training code, livestreamed the entire training run, and cited roughly $3.5M in total training cost, with the report breaking out $2.6M for Pro's RL stage and $0.9M for Flash. It was the most-watched open-source event of the day.

Confirmed

  • Two flagship models: MiMo-V2.6-Pro with 1.02T total / 42B active parameters, and MiMo-V2.6-Flash with 310B total / 15B active—both Sparse MoE, natively multimodal (text/image/video/audio), 1M-token context, MIT-licensed (cross-confirmed by @natolambert, @ArtificialAnlys and others)
  • Pro scored 46 on the AA Intelligence Index, the top open-weight model; predecessor MiMo-V2.5-Pro scored 26 (@ArtificialAnlys); per @heyshrutimishra citing AA rankings, Pro's weighted cost is about $0.13 per task
  • Architecture unchanged from V2.5, adding a 1M-token context window and an MTP drafter; early evaluations show scores just behind GPT-5.6 Sol (@mervenoyann)
  • Beyond weights, the tech report, RL environments and training code are open-sourced, with 7000 RL training environments planned; the official claim is that Pro matches Claude Opus 5 / GPT-5.6 Sol-class models on most agent benchmarks and entered the Code Arena top 10 (@arena, @teortaxesTex)
  • API pricing is unchanged from V2.5 for both Pro and Flash; officially claimed to set a new cost-performance record for Chinese models at 1/20 to 1/60 of international pricing (official line via @amaarora)
  • Per @wearemammals relaying official figures, total training cost was about $3.5M, positioned for extreme cost-effectiveness, with a live benchmarking dashboard publicly tracking evaluations; training was livestreamed (@Prompt Engineering)
  • The tech report details: $2.6M for Pro's RL training, $0.9M for Flash, with 44% of compute on rollouts (via @KyeGomezB); @andersonbcdefg relays AA's report that 45k RL tasks cost about $2.6M; @pcuenq notes per-task cost is a small fraction of peers, and the RL livestream has ended
  • Luo Fuli wrote a long post: amid extreme compute scarcity, Xiaomi committed a team of dozens to scaling up RL long-term, with a mid-training approach—likely the largest single RL run by compute among open-source teams, harder than DeepSeek R1 (relayed by @timshiai, @teortaxesTex)
  • @deedydas tested MiMo 2.6 Pro hands-on, found it quite good and willing to answer security-research questions, estimating agentic-coding pricing (with cache-hit assumptions) at 15x cheaper than Kimi K3, 6x cheaper than GLM 5.3, and about half of DeepSeek V4
  • @TheAhmadOsman (relayed by @Jasonio) noted Pro is only 1 point below GPT 5.6 Sol Max yet runs on 8x RTX PRO 6000
  • Reddit discussion (@power97992) regards MiMo 2.6 Pro as the strongest open model on aggregate benchmarks, lists Xiaomi, Z-AI and Kimi as the leading Chinese labs, but notes real-world usage still has misses; @arpitbhayani also called it much better than other models, surprised Xiaomi wasn't on his watchlist
  • The Hugging Face repos include MiMo-V2.6-Distill-Qwen-9B (image-text-to-text distilled) alongside Pro-RL and Flash-RL (@Bestlife73, @AdinaYakup); MiMo-V2.6-Flash-RL hit the Hugging Face trending list (@XiaomiMiMo)
  • @tokenbender observed from Flash-RL's config that it supports image, audio and video inputs with built-in tool-call rendering templates for agent workflows
  • Per @natolambert and @giffmana's discussion, the MiMo team released the model within about a day of finishing RL training

Unconfirmed

  • @mervenoyann noted AA benchmark scores were incomplete at release; @victormustar cited 524B parameters for Pro, inconsistent with most posts' 1.02T total—check the model page; the $3.5M total cost vs $2.6M (Pro RL stage) figures differ—given the report's $2.6M/$0.9M split, the former likely includes pretraining while the latter covers RL only

Why it matters

  • A phone maker's model topping the third-party intelligence index for open weights, with trillion-parameter weights and RL training code fully released, materially shifts the open-source landscape
  • The jump from 26 to 46 sets what @victormustar calls a new open-model record on the index; the RL scale described by Luo Fuli, the relatively low $3.5M total cost, the livestreamed training and 7000 planned open RL environments are seen by the community as a signal that large-scale RL is becoming reproducible by anyone
  • API pricing held at V2.5 levels, reinforced by @deedydas's hands-on cost estimates and the $0.13 per-task weighted cost, strengthening the cost-performance claim

27 more related posts →

Episode 3 · Xiaomi's MiMo-V2.6 Pro tested: big gains in coding, 3D games and multimodal (2026-09-22, 2 posts)

Hands-on tests of Xiaomi's open-source MiMo-V2.6 Pro cover coding, frontend, 3D game development and multimodal tasks, with Agentic Coding scores roughly tripling over the previous version and peak speeds hitting 900tps.

Episode 4 · Xiaomi Open-Sources MiMo-V2.6, Scoring 46 on AA Index, Just 1 Point Behind GPT-5.6 (2026-09-24, 7 posts)

On 09-24 Xiaomi released and open-sourced the MiMo-V2.6 family (Pro and Flash), two natively multimodal models built around scaled reinforcement learning. Pro scored 46 on the Artificial Analysis Intelligence Index, topping all open-weight and Chinese models, just 1 point behind GPT-5.6 Sol (max) at 47, while its weighted average cost is only about $0.13 per task—roughly 1/15 of GPT-5.6's. The gap between open-source and top closed frontier models narrowed to under one point with an order-of-magnitude cost advantage, with direct implications for model pricing and the open-source ecosystem.

Confirmed

  • MiMo-V2.6 includes Pro and Flash, both natively multimodal, emphasizing scaled reinforcement learning
  • Pro scored 46, first among open-weight and Chinese models, ahead of Qwen3.8-Max, Kimi K3, GLM-5.3 (@FellMentKE) and DeepSeek V4.1 Flash at 39 (@buffering112)
  • GPT-5.6 Sol (max) scored 47; Pro's weighted per-task cost is $0.13, about 1/15 of GPT-5.6's (@kimmonismus)
  • @buffering112 reports user-measured costs 15-20% lower than DeepSeek
  • Pro's agent-benchmark performance is said to be on par with Claude Opus 5 and GPT-5.6 Sol

Why it matters

  • @davidyin44 notes this continues the trend of Chinese vendors offering near-frontier closed-model capability at very low cost; Xiaomi, a phone maker, entering open-source models intensifies competition
  • A 1-point gap with roughly an order-of-magnitude cost difference could disrupt closed-model pricing logic and boost the practical value of open-source models