MiMo-V2.6: Xiaomi's Trillion-Parameter Open Model, Trained in Public

🎯 The Live-Dashboard Release 0:00

Somewhere on the internet right now is a live dashboard showing exactly how much money a major tech company is burning to train an AI model — "not a summary after the fact, not a polished blog post… a live number ticking upward in real time while engineers watch their training run succeed or fail in front of the entire world." That's what Xiaomi did. Out the other side came MiMo-V2.6: over a trillion parameters, a training bill north of $3 million shown openly, and a model anyone can download for free.

The framing that carries the whole story: a "strange and almost reckless level of transparency" behind how it was built, benchmark numbers that put it in the same conversation as the biggest closed models, and "a few details, if you look closely enough, that don't quite add up."

📈 The Training Run: $3.47M in Public 1:56

The MiMo team — led by Luo Fuli, previously at DeepSeek — had gone quiet for nearly six months, focused on one question: how far can reinforcement learning actually be scaled? Rather than disappear and drop a finished model with cherry-picked results, they let the entire process play out in public, "mistakes included" — and that part is literal. The dashboard showed reward curves, token throughput, and a cost counter climbing hour by hour, but also the failures as they happened: a memory fault that restarted a run, a network issue that dropped a whole category of training data, and an infrastructure error that went unnoticed for hours. Each training step processed 1,500 prompts with 16 attempts each, all in parallel.

Then, days of silence — the finished models sat marked "stopped" with no announcement. On September 21 at 15:39 UTC, the weights simply appeared on Hugging Face: Flash first, Pro 18 seconds later. No drama, no blog post — just two repositories suddenly live and free. The formal announcement, blog, and API came the next day.

👨‍👩‍👧 The Model Family 5:12

Xiaomi shipped a family, not a single model:

• MiMo-V2.6 Pro — the flagship: 1.02 trillion total parameters, ~42B active. Heavy research, long agent work, high-stakes coding.
• MiMo-V2.6 Flash — the practical everyday choice, trading a little power for speed and cost efficiency.
• MiMo-V2.6 Pro UltraSpeed — the same Pro model tuned to respond up to ~20× faster, for real-time interactions.
• MiMo-V2.6 Distill-Qwen-9B — a distilled model for people without massive server hardware.

Every model in the lineup is natively omnimodal — text, images, video, and audio processed by one unified model rather than bolted-together systems — and every one shares a 1M-token context window (enough for entire codebases, hours of audio, or very long documents).

🏗️ Architecture: MoE, Omnimodal, Speculative Decoding 6:44

The design is a sparse mixture of experts: Pro holds just over a trillion total parameters, but only ~42 billion are active at any moment — many small specialized sections, with only the relevant ones switching on per task. That's why a trillion-parameter model still runs at reasonable speed and cost.

Taking Flash as the example (Xiaomi's most detailed spec sheet): 48 layers, of which 39 use sliding-window attention (a narrow window that slides along) and 9 use full global attention (looking across the entire input) — a common mix for speed plus long-range awareness. Hidden size 4,096, 256 experts with 8 active per input, and no permanently-active shared experts. Feeding the backbone are three specialized components that all route into the same shared brain: a vision encoder (~681M params), an audio tokenizer (~308M), and a smaller audio patch encoder (~127M).

🖥️ Hardware Reality: 173GB, Multi-GPU 11:18

"Open source" here needs context. Flash alone ships as ~173GB across 65 file shards in a compressed format, and Xiaomi's own recommendation is serving frameworks like SGLang or vLLM spread across multiple GPUs. "In plain terms, this is not something you are casually running on a laptop, or even a single high-end gaming computer" — it's built for multi-machine server hardware. The license is free; the infrastructure isn't.

One more oddity: the auto-generated summary on Hugging Face lists Flash as 159B parameters — matching neither of Xiaomi's official figures — most likely a quirk of how the page counts parameters in a compressed format, but another thread left slightly untied.

📊 Benchmarks vs Opus 5 & GPT-5.6 Sol 12:41

On the Artificial Analysis Intelligence Index, Pro scores 46.32 — Xiaomi's claim is that it's the highest-scoring open-source model. The detail matters more than the headline:

Two patterns are worth sitting with. First, the model family is genuinely competitive — on autonomous-agent work (AutomationBench) Flash beats both frontier closed models, and on coding it sits within a couple of points. Second, security is uneven: excellent on CyberGym (a huge generational leap from V2.5's 40), but far behind on offensive exploit benchmarks — "specific strengths and weaknesses rather than a uniformly strong security skill set."

💰 Pricing & API Compatibility 17:15

Pricing is tiered by cached vs uncached input and output tokens, with Flash priced below Pro and UltraSpeed a clear premium for the speed (roughly an order of magnitude for the dramatically faster responses). The headline fact: Xiaomi confirmed the pricing is identical to the previous V2.5 generation — a meaningful capability jump at no extra cost, which is the opposite of how model pricing usually moves.

The practically useful detail is compatibility: the API is OpenAI- and Anthropic-format compatible, so switching an existing project "can be as simple as changing your base URL and your model name" — it slots directly into tools like Claude Code or Cline. There's also a token-plan subscription for high volume and a limited-time free tier ("MiMo Claw") for trying it out.

The ecosystem around it: MiMo Desktop (all-in-one pro app), MiMo Code (dedicated coding assistant), MiMo Studio (zero-setup chat), and a new Batch API for large-scale non-realtime workloads.

🧪 The Demos: Vibe World & Material Science 20:28

Two demos signal where Xiaomi wants to position this. Vibe World turns a text description, image, or video clip into a playable 3D environment — not in a single pass, but by coordinating multiple internal agents (one builds the physical scene, one handles object/character interactions, one visually checks the result). It's a small team of specialized processes working in sequence.

The more impressive one is internal and scientific: Xiaomi used MiMo-V2.6 Pro as a research assistant for its material-science team, helping design a metal-organic framework to capture "forever chemicals" (extremely persistent pollutants). The model reviewed existing literature, formed original hypotheses about which designs might work, and ran dry experiments — simulations of how strongly candidate materials would bind to the pollutant — before anything reached a physical lab. The through-line: Xiaomi is positioning MiMo-V2.6 as an agentic system that coordinates complex multi-step work, not another chatbot.

📜 Licensing: What "Open" Really Means 22:23

The weights for Pro and Flash are released under the MIT license — about as permissive as it gets: no revenue cap, no research-only restriction, no share-back requirement. Commercial use, modification, fine-tuning, redistribution — all free, plus a technical report and the smaller distilled model.

But there's a distinction worth flagging precisely. During the live broadcast, Luo Fuli said the team intended to also open-source the RL training code and the training environments — the actual recipe, not just the finished result. As of the recording, that piece hadn't shipped: you got the weights, the high-level report, and deployment instructions, but not the training code. (Subsequent coverage notes Xiaomi has since opened 7,000+ RL environments.) If you just want a powerful, freely-licensed model, everything is there now; if you want to study or replicate the large-scale RL process, that part lagged the weights.

⚖️ Verdict 24:15

What Xiaomi put out is genuinely significant: a trillion-parameter omnimodal flagship, a full family, MIT licensing, pricing frozen from the previous generation, and benchmarks shoulder-to-shoulder with — and occasionally ahead of — the frontier closed models. For something that costs nothing to license and is fully open to inspect, that's impressive.

The honest counterweight: every benchmark is self-reported and unverified, there's an unexplained gap between the live-dashboard score and the final card, a smaller model outscoring its bigger sibling on one benchmark, and the training recipe arrived after the weights. None of it erases what's real here — but "the full story is still being written" as independent testers get their hands on it. The right posture is the one the video itself recommends: treat the numbers as a strong opening claim, and wait for third parties to confirm.

💡 Key Takeaways

1. Radical transparency. Xiaomi streamed the RL training live — reward curves, cost counter, and the failures (memory fault, dropped data, unnoticed infra error).
2. A real family, MIT-licensed. Pro (1.02T / 42B active), Flash, UltraSpeed (~20× faster), and Distill-Qwen-9B — all natively omnimodal with a 1M-token context.
3. Sparse MoE is the efficiency engine. 1.02T total params but only ~42B active per token; Flash is the lean ~30.9B sibling.
4. Competitive on agent benchmarks. Flash beats Opus 5 and GPT-5.6 Sol on AutomationBench, and sits within a couple points on DeepSWE.
5. Security is uneven. CyberGym 95.1 (vs V2.5's 40) but far behind on exploit benchmarks — specific strengths, not uniform skill.
6. Every benchmark is self-reported. None independently verified; a ~2-point DeepSWE gap between the live dashboard and the final card is unexplained.
7. Same price as V2.5. A capability jump with no price increase, and the API is OpenAI/Anthropic-compatible — it drops into Claude Code or Cline.
8. "Open" ≠ "open everything." Weights and report shipped first; the RL training code and environments came later.

🔗 Resources & Links

• 📄 mimo.xiaomi.com/mimo-v2-6 — MiMo-V2.6 release note
• 🤗 huggingface.co/XiaomiMiMo — model weights (Pro, Flash, Distill-Qwen-9B)

Source video: youtube.com/watch?v=M7QeuzS35cM. Note: the video's "309B" for Flash is corrected to ~30.9B here, per Xiaomi's materials and independent coverage.

⏱️ Timestamp Index