MiMo-V2.6 Review: Xiaomi's Trillion-Parameter Open Model — and the Training Ledger (2026)

What Xiaomi Shipped

Xiaomi released and open-sourced the MiMo-V2.6 family on September 21–22, 2026, publishing weights to Hugging Face and ModelScope under the MIT license. Three models plus a distilled variant:

The multimodal stack is specific: a 681M-parameter MiMo ViT reads images and a 308M AudioTokenizer handles sound, all feeding one model rather than separate modality pipelines.

Weights are published as MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, and the Distill-Qwen-9B build, with SGLang or vLLM recommended for serving. Community GGUF and 4-bit MLX conversions appeared within hours.

The Headline Number

MiMo-V2.6-Pro scored 46 on Artificial Analysis's Intelligence Index — the highest of any open-weight model, ahead of GLM-5.3 at 45 and Kimi K3 at 44. Its predecessor, MiMo-V2.5-Pro, scored 26.

A twenty-point jump between generations is the kind of move that usually accompanies a change in training regime rather than a scale-up, and Xiaomi says that is exactly what happened. What makes it symbolically notable is the source: a phone manufacturer now tops the open-weight intelligence board, above the labs whose entire business is models.

The gap to closed frontier systems remains wide, though. Claude Fable 5.1 and GPT-6 Astra both score 53 — seven points ahead of the best open model.

Xiaomi's own benchmark table:

Third-party corroboration comes from Code Arena's WebDev board, where MiMo-V2.6-Pro measured 1628 (AutoEval) — roughly tenth overall and third among open-weight models, up 153 points from V2.5-Pro's 1475. A live arena measurement agreeing with the index is better evidence than either alone.

Treat the rest as vendor-reported until the community reruns them, particularly the CyberGym leap from 40 to 94, which is the largest single-generation move in the table.

The Part Nobody Else Does: The Training Ledger

Xiaomi's release is unusual for a different reason than benchmark scores. It streamed its reinforcement learning runs for six days and published the bill.

Both runs completed inside the week after roughly 750,000 trajectories each, for a combined bill near $3.47 million. Xiaomi reports relative pass-rate gains of about 25% for Flash and 12% for Pro on the training tasks, with DeepSWE v1.1 improving roughly 17 points for Flash (48.8 to 65.68) and 14 for Pro (58.4 to 72.57) during the RL run itself — a different measurement from the launch table above, which compares final models across generations. The training method is described as "You Only RL Once": fully asynchronous GRPO running 1,568 prompts per step with 16 rollouts each, plus groupwise agentic grading and multi-prefix multi-teacher on-policy distillation.

Why that number matters: labs almost never publish real post-training costs. $3.47M gives the industry a calibration point, and the calibration is that lifting an open frontier model one generation is a low-single-digit-millions exercise — not pocket change, and not the hundreds of millions that frontier pre-training consumes. For anyone budgeting their own post-training work, that is a more useful data point than any benchmark.

And it published the bad days

The genuinely rare part is the failure log. Straight from the dashboard:

• The Pro run restarted at step 17 after a GPU out-of-memory failure caused by expert load imbalance — "adjusted the training parallelism strategy"
• A network connectivity issue between the Pro training cluster and the grader deployment forced a restart; Xiaomi also removed the cyber dataset from the upcoming Pro run after observing "bad patterns in the rollout logs"
• Another Pro restart from a VRAM issue on one node
• The Flash run restarted from step 15 after an infrastructure error on one dataset went undetected for roughly three hours

A release that publishes its own OOM restarts and a three-hour silent failure is more credible than one that only ships good news, and the disclosure of pulling a dataset for bad rollout patterns is the kind of detail that usually stays internal. The caveat is that this is Xiaomi's own accounting with no third-party audit — it is a transparency claim, and the numbers are self-reported.

Pricing: Flat, Which Is the Point

API prices are unchanged from V2.5, per million tokens:

Flash's 15B active parameters are what makes that $0.14 input rate possible. A model that activates 15B per token at $0.14 per million input tokens sits with the cheapest credible options in the market — and as our DeepSeek V4.1-Flash review shows, the cache-hit column is often what decides the bill on agent workloads.

The strategic reading: capability up 20 index points, price unchanged. Xiaomi had already cut MiMo API prices by up to 99% in May 2026, and V2.5 took first place on OpenRouter by both monthly and weekly call volume. Holding the price while raising capability moves the value anchor for open models further left rather than harvesting the gain — which is what a vendor does when distribution matters more than margin.

Who Should Use It

Summary

MiMo-V2.6 is the strongest open-weight release of the quarter on the numbers, and the most interesting part is not the numbers. A phone maker shipped a 1.02T-parameter omni-modal model under MIT, topped the open-weight intelligence index, held prices flat, and published a $3.47 million training ledger complete with OOM restarts and a three-hour silent failure.

The honest limits are equally clear. Every capability figure is Xiaomi's own, the "20x faster" UltraSpeed claim has no measurement behind it, the frontier is still seven index points ahead, and Pro needs a GPU node rather than a desktop. The 9B distill is the pragmatic entry for local work.

Two things are worth watching. Whether the community reproduces the CyberGym and DeepSWE jumps — that is the claim carrying the most weight. And whether anyone else follows on disclosure: if publishing your RL bill and failure log becomes normal, the cost of overclaiming rises for everyone.

For the open-weight competition, see DeepSeek V4.1-Flash, Qwen3.8-Omni-Flash, and Kimi K3; for how to spend these models efficiently, the code review cost analysis.

Related Articles

• DeepSeek V4.1-Flash Review
• Qwen3.8-Omni-Flash Review
• Kimi K3 Review
• Best Chinese AI Models 2026