Xiaomi MiMo-V2.6-Pro: Top Open-Weights Model (2026)

Xiaomi released the MiMo-V2.6 series on 22 September 2026 (China time; the Hugging Face repositories went up late on 21 September UTC). It comes five months after MiMo-V2.5, and the jump is large. On Artificial Analysis's independent Intelligence Index v4.3.2, V2.5-Pro scores 25.99 and V2.6-Pro scores 46.32. That puts it ahead of every other downloadable-weights model AA tracks, including GLM-5.3 (Max) at 44.78, Kimi K3 (Max) at 43.59 and DeepSeek V4.1 Flash (Max) at 39.46. The weights are MIT, with no revenue thresholds or Model-as-a-Service carve-outs.

This guide covers the six checkpoints Xiaomi has published, the specs, what the benchmarks show and where they come from, what the API costs, what hardware you need to self-host, and how V2.6-Pro compares with the other open-weight flagships. It also covers the part many launch stories skip: twelve days before this release, Xiaomi was named in Anthropic's report on distillation of Claude.

What is MiMo-V2.6?

MiMo-V2.6 is a family of four models from Xiaomi's MiMo team, shipped across six Hugging Face repositories. Xiaomi's release notes frame the series around one idea: "scaling reinforcement learning toward self-improvement". In practice that meant one mixed RL run across coding, general agents, visual tasks and cybersecurity, where earlier releases used separate runs for each domain.

• MiMo-V2.6-Pro is the flagship. It is an omnimodal sparse MoE with 1.02T total and 42B active parameters, published on Hugging Face as XiaomiMiMo/MiMo-V2.6-Pro-RL.
• MiMo-V2.6-Flash is the cheaper sibling: 309B total and 15B active, the same omnimodal stack and a 1M context (XiaomiMiMo/MiMo-V2.6-Flash-RL).
• MiMo-V2.6-Pro-UltraSpeed is available only through the API. Xiaomi says it generates output "up to 20 times faster" than standard Pro, and charges 10x the price for it.
• MiMo-V2.6-Distill-Qwen-9B is a 9B checkpoint made by supervised fine-tuning of Qwen3.5-9B on data generated by MiMo. It is released as a research starting point for agentic RL. It is not a flagship.
• The MOPD checkpoints, published six days later on 27 September 2026, are the ones to pull for agent work: XiaomiMiMo/MiMo-V2.6-Pro-MOPD and XiaomiMiMo/MiMo-V2.6-Flash-MOPD. Same architecture, same 573 GB and 178 GB on disk, same MIT licence — plus a short multi-teacher on-policy-distillation pass that fixes a tool-call repetition failure mode where the model kept re-issuing near-identical tool calls and made no progress.

A naming trap worth knowing before you type a download command. There is no XiaomiMiMo/MiMo-V2.6-Pro repository and no XiaomiMiMo/MiMo-V2.6-Flash repository. Xiaomi only publishes the suffixed checkpoints — -Pro-RL, -Pro-MOPD, -Flash-RL, -Flash-MOPD and -Distill-Qwen-9B. The bare names are API model IDs, not Hugging Face repos, so any hf download or vllm serve line using them fails with a 401 rather than a clear "not found".

Pro and Flash take text, image, video and audio as input and produce text. Xiaomi also published the technical report and open-sourced the tooling around the release on the XiaomiMiMo GitHub org, which added a verl RL-framework fork, a long-horizon agent framework called Uni-Agent, and a minimal coding agent called mimoagent on 21 September.

MiMo-V2.6-Pro specs

These figures come from the Pro and Flash model cards on Hugging Face and the repositories' safetensors metadata.

Two design choices explain most of what the model is good at. Sixty of Pro's seventy layers use a 128-token sliding window, and only ten attend globally. That keeps the KV cache small enough to make a 1M context practical at serving time. The five-layer multi-token-prediction drafter drafts seven tokens per forward pass, which is how a 1.02T-parameter model stays servable at all — though it does not make it fast. Artificial Analysis measures Pro's median output at 41.18 tokens per second, with a 4.19-second time to first chunk. That is below most of the models it beats on score, so if interactive latency matters more to you than peak quality, that is the trade.

The checkpoint is smaller than 1.02T parameters would suggest. Most of the stored tensors are in a packed 8-bit container, with FP8 and BF16 used for the rest, so the download is about 573 GB instead of the roughly 2 TB a BF16 release would need. Xiaomi does not name the exact quantisation scheme on the model card.

How good is MiMo-V2.6-Pro? Benchmarks

There are two sets of numbers: Xiaomi's own evaluations and Artificial Analysis's independent index. Give the independent number more weight.

Independent: Artificial Analysis Intelligence Index

Artificial Analysis scores MiMo-V2.6-Pro at 46.32 on Intelligence Index v4.3.2, the highest figure on the board for a model whose weights you can download. The models directly behind it are Qwen3.8 Max (0902) at 45.42, GLM-5.3 (Max) at 44.78, Kimi K3 (Max) at 43.59, GLM-5.3-Flash at 41.81 and DeepSeek V4.1 Flash (Max) at 39.46. Two caveats on reading that list. AA tracks two dated Qwen3.8 Max snapshots — 40.15 and 45.42 — and Alibaba has not said which one corresponds to the released Qwen3.8-2.4T-A95B weights, so treat Qwen's position as a range. And AA's field has grown past 220 scored model-variants, so we have dropped the "#N of N" rank we previously quoted: a rank against an older field size is stale even when the score behind it is right. VentureBeat reported at launch that Pro ties Grok 4.7; on v4.3.2 that holds only against Grok's top effort tier (Grok 4.7 Xhigh, 46.45) — it beats the default Grok 4.7 tier, which scores 42.22. The Next Web notes that Claude Opus 5.5 leads the overall board; on v4.3.2 it scores 57.62. MiMo leads the open-weight field. It does not lead the whole field.

A note on reading these numbers: every score on this page is Intelligence Index v4.3.2, captured 5 October 2026. AA rescaled the index from v4.1.1 when it swapped in harder evaluations such as Terminal-Bench 4.0 and AutomationBench, and there is no conversion factor between versions — on a rebase every score moves because the benchmark basket changed, not because the models did. Scores published in August under v4.1.1 (for example, Kimi K3 at 60 in our Kimi K3 guide) are not comparable with today's numbers. Compare models within one index version only.

Cost is the other half of AA's verdict. AA measures Pro at $0.1332 per Intelligence Index task, which puts it on AA's intelligence-versus-cost Pareto frontier, and at $206.66 to run its whole v4.3.2 suite. Per-task is the figure to compare across models — AA runs different task counts per model, so suite totals are each model's own absolute cost, not a ratio you can divide. On per-task cost, Pro runs the index for about a fifteenth of what Kimi K3 (Max) costs at $2.00. Pro is also verbose: it emitted 140M output tokens across the evaluation suite. Low per-token prices partly offset that, but budget for long reasoning traces.

Vendor-reported: Xiaomi's evaluation table

These are Xiaomi's numbers from the Pro model card, measured in Xiaomi's harnesses. Rows marked "MiMo" are internal benchmarks that nobody outside Xiaomi can reproduce.

Three things stand out. First, the gain over V2.5-Pro is very large on agentic benchmarks. DeepSWE goes from 19.0 to 71.9, and Terminal Bench 4.0 from 1.5 to 34.9. That fits Xiaomi's claim that the RL recipe, not a new base, is what changed. Second, Pro is close to Opus 5 and GPT-5.6 Sol on older agent suites such as Terminal Bench 2.1 and OSWorld-Verified. It clearly trails on the hardest current ones: Terminal Bench 4.0 (34.9 vs 49.0) and ExploitBench (47.9 vs 70.0–78.5). Third, Flash is almost as strong as Pro on coding and general-agent tasks, and the gap only opens on offensive-security work.

Xiaomi's own release-notes page gives slightly different DeepSWE figures from the model card (72.6 for Pro, 65.7 for Flash), which suggests different run settings. We have quoted the model card throughout.

How much does MiMo-V2.6 cost?

Xiaomi serves the models on its MiMo API platform, and they are also listed on OpenRouter. Xiaomi says pricing is unchanged from the V2.5 series. These are the list prices per million tokens, checked against OpenRouter's model catalogue on 22 September 2026:

Standard Pro is extremely cheap for a model at the top of the open-weight index. It costs about one-seventh of Kimi K3 on input and one-seventeenth on output. Cached input at $0.0036 per million makes long-running agents with a stable system prompt and repository context close to free on the input side. UltraSpeed is the exception: you pay 10x for lower latency, which is worth it only for interactive products where output speed is the bottleneck.

A minimal call through OpenRouter's OpenAI-compatible endpoint looks like this:

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_OPENROUTER_KEY",
)

resp = client.chat.completions.create(
    model="xiaomi/mimo-v2.6-pro",
    messages=[{"role": "user", "content": "Refactor this function to be async: ..."}],
    temperature=1.0,
    top_p=0.95,
)
print(resp.choices[0].message.content)

Xiaomi recommends temperature=1.0 and top_p=0.95. On Xiaomi's own platform, the IDs are mimo-v2.6-pro, mimo-v2.6-flash and mimo-v2.6-pro-ultraspeed. The models are also available in Xiaomi MiMo Studio, MiMo Code and the MiMo Desktop app.

How do you run MiMo-V2.6 yourself?

The MIT licence means you can self-host any of the four models commercially without signing anything. What you can realistically run depends on which model you pick.

MiMo-V2.6-Pro: datacentre hardware only

You need about 573 GB of weights in GPU memory before any KV cache. Xiaomi's reference configurations are:

• vLLM: one 8-GPU node with tensor parallelism 8. Weights alone need about 72 GB per GPU, so in practice that means 8x H200 (141 GB) or 8x B200-class cards. 8x H100 80 GB leaves almost no room for KV cache.
• SGLang: Xiaomi's tuned recipe uses two nodes (TP 16, DP 2, expert parallelism 16, DeepEP all-to-all, EAGLE multi-layer speculative decoding). That is the configuration to copy for production throughput.

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

Stable vLLM releases may lag behind. Xiaomi points to the MiMo recipe and a pre-built image (vllm/vllm-openai:mimov25-cu129) from the V2.5 cycle. For SGLang, use lmsysorg/sglang:latest and the SGLang MiMo cookbook. Either way, pass the mimo reasoning and tool-call parsers, or tool calling will not parse correctly.

MiMo-V2.6-Flash: one server, or a large workstation

Flash's checkpoint is about 178 GB, which makes it the practical self-hosting pick. Xiaomi's SGLang recipe uses TP 8 on one node. Community quantisations appeared within hours of release, including NVFP4 builds for Blackwell GPUs, a W4A8 build aimed at two RTX PRO 6000 cards, MLX 4-bit and mxfp4 builds for Apple Silicon, and early GGUFs. They are unofficial, so benchmark them on your own tasks before you trust them. On Apple Silicon, check that the quant's file size fits in your unified memory with room for context. Our self-hosting LLMs guide covers how to size memory and pick a serving stack.

MiMo-V2.6-Distill-Qwen-9B: laptop-class

The 9B distill is 18.8 GB in BF16 and has the most community quants: GGUFs from ggml-org and bartowski, plus MLX builds. A Q4 GGUF runs on a single 8–12 GB GPU or an M-series Mac:

ollama run hf.co/bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF:Q4_K_M

Set your expectations: this is an SFT checkpoint released for research, not a small V2.6-Pro. On Xiaomi's numbers it clearly beats its Qwen3.5-9B base on agentic tasks (Terminal Bench 2.1 from 27.0 to 37.1, SWE-bench Pro from 32.0 to 44.6), but it is a different class of model from Pro and Flash. If you want a strong local model that fits in 16–24 GB, compare it with Qwen3.8-27B before committing.

MiMo-V2.6-Pro vs Kimi K3 vs Qwen 3.8 Max vs DeepSeek V4.1 Flash vs GLM-5.3

Here are the open-weight flagships you would realistically shortlist in September 2026. Intelligence Index scores all come from the same Artificial Analysis index version (v4.3.2, captured 5 October 2026), and each is labelled with the effort tier AA scored, because the gap between a model's default and top tier is often larger than the gap between two models. Prices are first-party list prices per million tokens.

How to read it:

• Top open-weight capability at the lowest flagship price: MiMo-V2.6-Pro. It leads GLM-5.3 (Max) by 1.54 index points, which is close to noise, but it is about 3x cheaper than GLM-5.3 on input and 5x cheaper on output — and 15x cheaper per Index task. It also has the cleanest licence of the five.
• Cheapest per token with a strong score: DeepSeek V4.1 Flash. It is 6.86 index points behind, but its off-peak price undercuts even MiMo Pro, and its tiny active-parameter count keeps self-hosting cheap.
• Largest model and strongest ecosystem: Kimi K3 is still the biggest open-weight model and has broad provider coverage, but at 43.59 against MiMo's 46.32, its $3/$15 pricing is now hard to justify on index score alone. Our Kimi K3 benchmark breakdown has the task-level detail.
• Algorithmic coding: Qwen3.8-Max ranks well on LMArena Code, and on AA's newer 0902 snapshot (45.42) it is the closest thing to a challenger in this table — but its older snapshot scores 40.15, and its headline agentic claims have not held up under neutral harnesses.
• Coding with a cyber focus: GLM-5.3 (Max) is 1.54 points behind MiMo on the index at 44.78. Check its licence if you run a large inference business.

For the full landscape, including smaller models you can run on one GPU, see our open-source LLMs pillar guide.

Is MiMo-V2.6 the best open-source LLM in September 2026?

On the best independent aggregate available, yes. MiMo-V2.6-Pro has the highest Artificial Analysis Intelligence Index v4.3.2 score of any downloadable-weights model as of 5 October 2026 — 46.32 — and at $0.1332 per Index task it is by a wide margin the cheapest of the leading group to run. Three caveats apply:

• The lead is narrow. 46.32 against Qwen3.8 Max (0902) at 45.42 and GLM-5.3 (Max) at 44.78 is close to a tie; Kimi K3 (Max) at 43.59 is two points further back. On your own workload, prompt formats, tool schemas and harness details will matter more than one or two index points.
• "Open weights" is not "open source". The weights, report and RL tooling are public. The pre-training data is not.
• The frontier gap is still about 11 points. Closed flagships such as Opus 5.5 (57.62) still lead on the hardest agentic and security tasks, as the Terminal Bench 4.0 and ExploitBench rows above show.

What did Anthropic's distillation report say about Xiaomi?

On 10 September 2026, Anthropic published a threat intelligence report covering misuse of Claude between December 2025 and August 2026. One section, on "illicit distillation", names seven China-based labs: Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax. Anthropic alleges they used Claude outputs to train their own models, across roughly 190 million exchanges in total.

According to The Hacker News' reading of the report, the Xiaomi-attributed campaign (tracked as GTG-16008) ran for about 20 days in March and April 2026. It totalled more than 400,000 exchanges, which is the smallest of the campaigns with published counts. By comparison, Anthropic attributes more than 151 million exchanges to Alibaba. Anthropic alleges that Xiaomi "replayed user conversations and coding sessions from its own MiMo models to Claude, through OpenClaw and OpenCode coding harnesses, to bolster training data used for future models."

These are Anthropic's allegations, not findings by a court or regulator. As of publication, The Next Web reports that Xiaomi has not responded. Better Stack's summary reports that China's Commerce Ministry rejected the allegations as having "no factual or legal basis". The reported campaign dates (March–April 2026) fall before the MiMo-V2.5 release in late April. The coverage we reviewed does not tie the alleged exchanges to any specific MiMo model, V2.6 included.

For teams choosing a model, the practical questions are narrower. The MIT licence is Xiaomi's to grant, and nothing in the report changes it. If your organisation has procurement or policy rules about vendors named in allegations like these, or about models from particular jurisdictions, apply them the same way you would to Kimi, Qwen, GLM or DeepSeek. All four of those labs are named in the same report.

Should you use MiMo-V2.6?

• Use Pro via API if you want the strongest open-weight model at a price low enough to run agent loops without watching the meter. Test it against your current model on your own tasks first. It is verbose, and vendor benchmarks are not your workload.
• Use Flash for high-volume agentic or multimodal work. It is within a few points of Pro on most of Xiaomi's agent benchmarks at about a third of the price, and it is realistic to self-host.
• Self-host Pro only if you already run 8x H200-class nodes and need data residency or MIT-licensed weights you control.
• Skip it if your work depends on the hardest security or terminal-agent tasks. The closed frontier still leads clearly there.

If your team is building products on open-weight models like MiMo, Codersera can help you hire vetted remote developers who have shipped LLM infrastructure and agent systems.

FAQ

When was MiMo-V2.6 released?

Xiaomi announced the MiMo-V2.6 series on 22 September 2026 (UTC+8). The Hugging Face repositories went live late on 21 September UTC, which is why some trackers, including Artificial Analysis, list 21 September.

Is MiMo-V2.6-Pro open source?

It is open-weight under the MIT licence, which allows commercial use, modification and redistribution with no revenue thresholds. Xiaomi also published the technical report and RL tooling. The pre-training data is not public.

How many parameters does MiMo-V2.6-Pro have?

1.02 trillion total, with 42 billion active per token: 384 routed experts, 8 of them active. MiMo-V2.6-Flash has 309B total and 15B active.

How much does the MiMo-V2.6-Pro API cost?

$0.435 per million input tokens, $0.87 per million output tokens and $0.0036 per million cached input tokens. Flash is $0.14/$0.28, and Pro-UltraSpeed is $4.35/$8.70.

Which MiMo-V2.6 checkpoint should I download?

For agent and coding work, the MOPD checkpoints published on 27 September 2026 — XiaomiMiMo/MiMo-V2.6-Pro-MOPD or XiaomiMiMo/MiMo-V2.6-Flash-MOPD. They are the same architecture and size as the launch -RL checkpoints, with a distillation pass that removes a repeated-tool-call failure mode. Note there is no bare MiMo-V2.6-Pro repository on Hugging Face — only the suffixed ones — so a download command using the plain name will fail.

Can I run MiMo-V2.6 locally?

Pro needs about 573 GB of GPU memory for weights alone, so a single 8x H200-class node. Flash (about 178 GB) fits on one server, or on a large workstation with community 4-bit quants. The 9B Distill-Qwen model runs on a single consumer GPU or a Mac through GGUF or MLX builds.

Is MiMo-V2.6-Pro better than Kimi K3?

On Intelligence Index v4.3.2 it scores 46.32 to Kimi K3 (Max)'s 43.59, a lead of 2.73 points. Its API is roughly 7x cheaper on input and 17x cheaper on output, and AA measures it at $0.1332 per Index task against K3's $2.00. K3 is a larger model with more third-party hosting options.

Sources

• Xiaomi MiMo-V2.6 release notes
• MiMo-V2.6-Pro-RL model card (Hugging Face) · MiMo-V2.6-Flash-RL · MiMo-V2.6-Distill-Qwen-9B
• Artificial Analysis: MiMo-V2.6-Pro · AA open-weight leaderboard
• OpenRouter: MiMo-V2.6-Pro
• VentureBeat launch coverage
• Anthropic threat intelligence report, September 2026 · The Hacker News summary