跳到主要内容

MiMo-V2.6: Xiaomi's 1T Open-Weight Omnimodal Model

进阶

On September 22, 2026, Xiaomi published MiMo-V2.6 — three checkpoints on Hugging Face under a plain MIT license, headlined by a 1.02-trillion-parameter sparse Mixture-of-Experts model with a 1-million-token context that takes text, image, video and audio as input. It landed the same day as Claude Opus 5.5 and GPT-6 Sol/Luna, and it went straight to the top of Hacker News.

The headline write-ups all say roughly the same thing: frontier-class open weights, twenty to sixty times cheaper than closed models. That part is true and easy to check. What nobody has written up is the part that decides whether you should actually use it — which benchmark versions the wins came from, why the flagship is not a machine you own, and what the third price tier is for.

What you'll learn
  • Read the MiMo-V2.6 benchmark table the way its own model card reports it — including the 14-point Terminal-Bench 4.0 deficit the press coverage did not quote
  • Know why a 33-point generational jump on an agentic benchmark usually means the old model could not drive the harness, not that the new one got smarter
  • Understand the architecture that buys the 1M context: 384 experts with 8 active, and a 6:1 sliding-window-to-global attention ratio
  • Pick between Pro, Flash and the 9B distill by what hardware you actually have — Pro's own launch command wants two nodes
  • See what the 10x-priced UltraSpeed tier sells you (latency, from the same weights) and when that is worth paying for

The one-sentence version​

MiMo-V2.6-Pro is an MIT-licensed, 1.02T-total / 42B-active sparse MoE with a 1M-token context and text, image, video and audio input that ties or beats Claude Opus 5 on short-horizon agentic and automation benchmarks, trails it badly on long-horizon terminal work, and costs roughly one-tenth as much per token when you rent it instead of running it.

Four things that aren't obvious from the headlines​

1. The Terminal-Bench win and the Terminal-Bench loss are the same model​

This is the one to internalise. Launch coverage widely reported MiMo-V2.6-Pro at 89.9 on Terminal Bench 2.1, just ahead of Claude Opus 5's 89.1. Xiaomi's own model card does not report Terminal Bench 2.1 at all. It reports Terminal Bench 4.0, where MiMo-V2.6-Pro scores 34.9 against Opus 5's 49.0 — a 14.1-point deficit.

Both numbers are real. Terminal Bench 2.1 is an older, largely saturated harness where every frontier model clusters near 90 and a 0.8-point gap is noise. Terminal Bench 4.0 is the current one, and it is where models are still spread out. If you read "beats Claude on Terminal Bench" anywhere, the first question is always which version — because on agentic suites the version change is usually a bigger effect than the model change.

The same pattern shows up in AILmanac's coverage of DeepSeek-V4.1-Flash and Gemini 3.8 Flash: parity on DeepSWE, a 15–20 point hole on Terminal-Bench 4.0. It is becoming the signature shape of the cheap tier.

2. A 33-point generational jump is a harness story, not an intelligence story​

The model card's own table includes the previous generation. MiMo-V2.5-Pro scored 19.0 on DeepSWE v1.1 and 1.5 on Terminal Bench 4.0. V2.6-Pro scores 71.9 and 34.9.

A model does not get 23× better at terminal work in one generation. A score of 1.5 means the old model essentially could not complete the benchmark's tool-call protocol — it was failing at the harness, not at the reasoning. Read enormous generational deltas on agentic benchmarks as "we fixed tool calling" first, and only then as capability. The deltas on non-agentic benchmarks are the honest ones: ProgramBench went 12.5 → 26.5, roughly a doubling, which is a large but believable generational step.

3. The 1M context is bought with attention layout, not brute force​

MiMo-V2.6-Pro's LLM backbone is 70 layers: 60 sliding-window attention, 10 global attention — a 6:1 ratio. Sliding-window layers only keep a local KV window, so the per-token KV footprint at 1M tokens is a fraction of what a fully global 70-layer stack would need. The ten global layers are what still let information travel end to end.

The MoE is 384 routed experts with 8 active per token, giving 42B active out of 1.02T — about 4.1% of the model runs on any given token. Flash uses the same shape at 309–311B total and ~15B active. This is why a trillion-parameter model can be served at all: you pay 1T in memory and 42B in compute.

On top of that sits a 5-layer multi-token-prediction speculative decoder that drafts 7 tokens per pass. That is a throughput mechanism, and it matters for the next point.

4. "Omnimodal" means omnimodal in, and UltraSpeed sells latency, not quality​

Two things the marketing language blurs:

Modalities are input-side. The card lists text, image, video and audio as modalities, with a 681M-parameter MiMo ViT vision encoder and a 308M AudioTokenizer plus a 127M audio patch encoder. The Hugging Face pipeline tag for Pro is text generation. You can feed it a video and ask questions; it answers in text. It is not a speech or image generator. If you want generation across media, that is a different page: generative media.

UltraSpeed is the same model at 10× the price. MiMo-V2.6-Pro-UltraSpeed is described as Pro quality at up to 20× the output speed, listed at roughly ten times Pro's per-token price. There is no separate set of weights on Hugging Face for it — it is a serving tier built on the speculative decoder and Xiaomi's infrastructure. That is a perfectly reasonable product, but price it as what it is: you are buying tokens-per-second for interactive and real-time workloads, not accuracy. If your job is a background batch, UltraSpeed is a 10× overpay.

The benchmark table, as the vendor published it​

All figures below are Xiaomi's own runs, from the MiMo-V2.6-Pro-RL model card. No independent replication existed at the time of writing.

BenchmarkMiMo-V2.6 ProMiMo-V2.6 FlashMiMo-V2.5 ProClaude Opus 5GPT-5.6 SolClaude Fable 5
DeepSWE v1.171.967.919.074.073.070.0
ProgramBench26.526.012.537.025.033.0
MiMo Code Bench63.261.240.468.659.3—
AutomationBench v1.0.653.152.316.050.345.846.2
Toolathlon-Verified76.973.649.180.674.977.9
Terminal Bench 4.034.928.81.549.039.942.4
OSWorld-Verified82.080.8—83.483.086.0
CyberGym94.095.140.0———
MiMo VisualCoding72.371.5—70.073.469.1

Three readings worth having:

  • Flash is within a few points of Pro almost everywhere. 67.9 vs 71.9 on DeepSWE, 52.3 vs 53.1 on AutomationBench, and it beats Pro on CyberGym. At roughly a third of Pro's hosted price and a third of the parameters, Flash is the default unless you have measured a reason otherwise.
  • The competitor columns for CyberGym are empty. Xiaomi reported 94.0 and 95.1 and published no Claude or GPT comparison. Treat an unopposed number as unopposed — it tells you the model is competent at the task, not that it leads.
  • Pro wins exactly one row outright: AutomationBench. Everything else is a near-tie or a loss. "Frontier-class open weights" is fair; "beats Claude" is not.

On the aggregate side, Artificial Analysis' Intelligence Index v4.3 placed MiMo-V2.6-Pro at roughly 46, reported as first among open-weight models and level with Grok 4.7, against about 53 for the closed frontier. A seven-point index gap for a tenth of the price is the actual pitch.

What it costs, rented and owned​

Rented, as reported at launch from the OpenRouter and Xiaomi listings, per million tokens:

ModelInputOutput
MiMo-V2.6-Flash$0.14$0.28
MiMo-V2.6-Pro$0.435$0.87
MiMo-V2.6-Pro-UltraSpeed$4.35$8.70

For scale, Claude Opus 5.5 lists at $4 / $20 (see the Opus 5.5 field guide). Pro is roughly 9× cheaper on input and 23× cheaper on output. UltraSpeed costs about the same as Opus 5.5 on input and less than half on output — which is the honest way to think about that tier.

Owned is a different conversation, and the launch commands give it away. The model card's SGLang example for Pro asks for --tp 16 --dp 2 ... --nnodes 2: tensor parallelism across sixteen devices, two data-parallel replicas, two nodes. That is a rack, not a workstation. The vLLM path is lighter at --tensor-parallel-size 8, still eight accelerators for a 1.02T checkpoint in BF16.

Flash, at ~309–311B total and ~15B active, is the variant that a serious single-node box can plausibly serve. The MiMo-V2.6-Distill-Qwen-9B checkpoint — a 9B distillation, listed on Hugging Face under image-text-to-text — is the one that fits on a laptop-class GPU. If your goal is "run a capable open model at home", start there and at Run models locally with Ollama, not at Pro.

One more ownership caveat: the repositories published are the post-RL checkpoints (-Pro-RL, -Flash-RL). No base checkpoint was released. You can serve, quantise and adapt these, but you cannot start your own post-training from the pre-RL base.

How to evaluate it for your own work​

Guided walkthrough1 of 5
  1. If your workload never sends an image, video or audio clip, most of what makes MiMo-V2.6 distinctive is irrelevant to you and a cheaper text-only open model may match it. The omnimodal encoders are the differentiator — use them or discount them.

Call MiMo-V2.6-Flash through an OpenAI-compatible endpoint

curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
  "model": "xiaomi/mimo-v2.6-flash",
  "messages": [
    {"role": "user", "content": "Summarise the failure modes in this log and propose one fix."}
  ]
}'

Serve MiMo-V2.6-Pro with vLLM (the model card's own launch line)

vllm serve XiaomiMiMo/MiMo-V2.6-Pro-RL \
--tensor-parallel-size 8 \
--trust-remote-code \
--gpu-memory-utilization 0.95 \
--max-model-len auto \
--reasoning-parser mimo \
--tool-call-parser mimo \
--enable-auto-tool-choice \
--generation-config vllm
Pro tip
  • Note the two parser flags: --reasoning-parser mimo and --tool-call-parser mimo. MiMo emits its reasoning and tool calls in a family-specific format. Omit these and your serving stack will hand the raw markers back to your application as ordinary text — a silent failure that looks like the model ignoring your tools.
  • --trust-remote-code is required because the modelling code ships with the checkpoint. Read it before you run it on anything you care about; that flag executes vendor Python in your serving process.
  • The SGLang path adds EAGLE speculative-decoding flags. That is where the throughput advantage lives — a vanilla launch leaves the multi-token predictor unused.

When to reach for it, and when not​

Good fit

  • High-volume classification, extraction or automation over mixed media, where $0.14/M input changes what is economically possible at all.
  • Anything you must run on your own hardware for privacy or jurisdiction reasons and that needs vision or audio understanding — few MIT-licensed models offer that combination. See a private local AI stack.
  • Short, bounded agentic tasks with well-specified tools, which is where AutomationBench and Toolathlon scores say it is genuinely competitive.

Poor fit

  • Long autonomous coding sessions. The Terminal Bench 4.0 gap is 14 points against Opus 5 and 5 points against GPT-5.6 Sol, and it is the vendor's own number.
  • Anywhere you need a base checkpoint to post-train from. Only RL checkpoints shipped.
  • Single-box self-hosting of Pro. Budget for Flash or the 9B distill, or rent Pro.

What Xiaomi disclosed that most labs do not​

Worth noting because it is unusually specific: Xiaomi published the RL training run's shape and cost — roughly a six-day run, 30 RL steps, about 750,000 trajectories, 1,568 prompts with 16 rollouts per step, at a reported ~$2.62M for Pro and ~$850K for Flash, using asynchronous GRPO with the MoE router frozen during RL.

Two things to take from it. First, a frozen router during reinforcement learning is a stability choice: it keeps expert assignment fixed so the policy update cannot quietly reshape which experts fire. Second, and more useful: a frontier-adjacent post-training run now costs single-digit millions, not hundreds. That is the number that explains why four labs shipped in the same week.

Check yourself

0/5
  1. Press coverage said MiMo-V2.6-Pro beat Claude Opus 5 on Terminal Bench; the model card shows it 14 points behind. What explains the contradiction?
  2. MiMo-V2.5-Pro scored 1.5 on Terminal Bench 4.0; V2.6-Pro scores 34.9. What is the most likely reading?
  3. What does the MiMo-V2.6-Pro-UltraSpeed tier actually sell at roughly 10x Pro's price?
  4. Why does MiMo-V2.6-Pro's 1M-token context not require a proportionally enormous KV cache?
  5. You want to run a capable open-weight model with vision on one workstation GPU. Which MiMo-V2.6 checkpoint is the realistic starting point?
按 Enter 或空格键翻转卡片。使用左右方向键在卡片之间切换。已显示术语。
1 / 9
Key takeaways
  • Check the benchmark version before the benchmark score. MiMo-V2.6-Pro beats Opus 5 on Terminal Bench 2.1 and trails it by 14 points on 4.0 — same model, same vendor, opposite headline.
  • Flash, not Pro, is the default. Within a few points on every published row, ahead on CyberGym, roughly a third of the price, and actually servable on one node.
  • Omnimodal is input-side. Video and audio go in, text comes out. If you never send media, a cheaper text-only open model may do the same job.
  • UltraSpeed is a latency product at 10x the price of identical quality. Buy it for interactive work; never for a batch.
  • The 1M context rides on 10 global attention layers out of 70. Test long-range retrieval at your real context length before trusting the ceiling.
  • Only RL checkpoints shipped, and every benchmark is a vendor run. Build a 20-task private eval in the shape of your own work before committing.

Sources & further reading​

Next​