MiMo-V2.5-Pro

Overview

MiMo-V2.5-Pro is Xiaomi's open-weight flagship large language model, released on April 22, 2026, with open weights published to Hugging Face five days later on April 27, 2026. Xiaomi's model card describes it as "an open-source Mixture-of-Experts (MoE) language model with 1.02T total parameters and 42B active parameters," using "the hybrid attention architecture and 3-layers Multi-Token Prediction (MTP) introduced in MiMo-V2-Flash, with up to 1M tokens context length."

The headline is the license. A consumer-electronics manufacturer has published a trillion-parameter model under MIT — an OSI-approved license with no field-of-use clause, no geographic carve-out, and no acceptable-use addendum. Xiaomi's release note frames it as permitting "free commercial use, secondary training, and fine-tuning without additional authorization." The Hugging Face API returns license:mit, and arena.ai's leaderboard metadata independently records the license as MIT. It is not the largest permissively licensed model, though — Artificial Analysis's dataset records DeepSeek V4 Pro (1.6T) and Ling-2.6-1T (1.03T) as MIT as well.

The pitch is agentic endurance at low token cost. Xiaomi's launch post claims that "when paired with a proper harness, V2.5-Pro can sustain complex, long-horizon tasks spanning more than a thousand tool calls," and its product page asserts "Peak Agent Performance: Rivals Claude Opus 4.6 in demanding agentic workloads." Xiaomi's release note further claims the model "ranks first among open-source models globally on the GDPVal-AA and ClawEval leaderboards." These are vendor claims; the independent numbers below are more mixed.

A note on naming

MiMo is no longer a 7B model. Xiaomi's first MiMo release, XiaomiMiMo/MiMo-7B-RL, landed on Hugging Face on April 29, 2025 with 7,833,409,536 parameters, and was positioned as a small reasoning model. Much of the writing still indexed under "Xiaomi MiMo" describes that model.

The current flagship shares only the brand. MiMo-V2.5-Pro reports 1,023,244,718,976 total parameters in its Hugging Face safetensors index — about 131× the 7B model. Two further sources of confusion:

• MiMo-V2.5 (no "-Pro") is a different, smaller, omnimodal model: 310B total / 15B active, with vision and audio encoders. It is not a variant of the Pro.
• The MiMo-V2 series was deprecated on June 30, 2026. Xiaomi's site states the V2 models "have been deprecated on June 30. Please migrate to the V2.5 series soon." Benchmarks or token-share statistics attributed to "MiMo-V2-Pro" describe a retired model.

Capabilities

• Long-horizon agentic execution: Xiaomi reports the model building a complete SysY compiler in Rust — "finished in 4.3 hours across 672 tool calls, scoring a perfect 233/233 against the course's hidden test suite" — and an 8,192-line desktop video editor "over 1,868 tool calls across 11.5 hours of autonomous work."
• Token efficiency: "On ClawEval, V2.5-Pro lands at 64% Pass^3 using only ~70K tokens per trajectory — roughly 40–60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable capability levels."
• Hybrid attention for cheap long context: The model card states the SWA/GA interleave "reduces KV-cache storage by nearly 7x while maintaining long-context performance via learnable attention sink bias."
• Multi-Token Prediction: "Equipped with three lightweight MTP modules using dense FFNs. This triples output speed during inference and will be good to accelerate rollout in RL training."
• Genuine 1M-token retrieval: On OpenAI's GraphWalks benchmark, Xiaomi reports that where its predecessor "collapses to 0.00 at 1M on both subtasks," V2.5-Pro "still scores 0.56 BFS / 0.92 Parents at 512k and 0.37 / 0.62 at 1M."
• Reasoning by default: The chat template enables <think> blocks unless enable_thinking is set false. Artificial Analysis classifies it as a reasoning model.
• Native tool-call format: A custom XML <tool_call> grammar, with dedicated mimo reasoning and tool-call parsers in SGLang.

Technical Specifications

• Hugging Face repository: XiaomiMiMo/MiMo-V2.5-Pro (base model: XiaomiMiMo/MiMo-V2.5-Pro-Base)
• Total parameters: 1,023,244,718,976 (1.02T), read from the safetensors index
• Active parameters: 42B per token
• Layers: 70 — 1 dense + 69 MoE
• Experts: 384 routed, top-8 activated, no shared experts
• Attention: Hybrid — 60 Sliding Window Attention layers + 10 Global Attention layers (6:1), 128-token window; 128 query heads, 8 KV heads (GQA); head dim 192 (QK) / 128 (V)
• MTP layers: 3
• Architecture: Mixture-of-Experts transformer, model_type: mimo_v2, custom modelling code (trust_remote_code)
• Precision: FP8 (E4M3) mixed — roughly 1.03 TB of weights on disk
• Context window: 1M tokens (max_position_embeddings: 1048576); the Base checkpoint is 256K
• Vocabulary: 152,576
• Pre-training: "Trained on 27T tokens using FP8 mixed precision and native 32k seq length"
• Post-training: SFT, large-scale agentic reinforcement learning, and Multi-Teacher On-Policy Distillation (MOPD)
• License: MIT

Xiaomi does not publish a knowledge cutoff for MiMo-V2.5-Pro — Artificial Analysis records the field as null — so this page does not state one.

Use Cases

• Autonomous coding agents: The model is tuned for harnesses like Claude Code, OpenCode, and Kilo; Xiaomi's own examples run for hours across hundreds of tool calls.
• High-token agentic workloads: At $0.435 in / $0.87 out, trajectories that burn tens of millions of tokens become affordable. Pareekh Jain, quoted by InfoWorld, frames the model as a "cost-efficient agent model for high-token workloads."
• Whole-repository and long-document reasoning: A 1M-token context window with measured 1M-token retrieval, rather than a nominal one.
• Self-hosted commercial deployment: MIT permits redistribution and commercial inference without a separate agreement — relevant where a restricted community license is a legal blocker.
• Distillation and continued pre-training: The Base checkpoint is published under the same license, making knowledge distillation and domain fine-tuning legally straightforward.
• Latency-tolerant batch work: Artificial Analysis measures a median time to first answer token of 48.2 seconds — the model thinks at length before answering.

Performance / Benchmarks

Independently measured (Artificial Analysis)

Artificial Analysis runs its own evaluations. Its Intelligence Index v4.1 combines nine benchmarks: "GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR."

For scale, Artificial Analysis's own comparison chart places MiniMax-M3 at 44.4, DeepSeek V4 Pro at 44.3, Kimi K2.6 at 44.2, MiMo-V2.5-Pro at 42.2, and GLM-5.2 at 51.1. MiMo-V2.5-Pro is competitive with, not ahead of, the open-weight field — while costing considerably less per token.

Vendor-reported

Xiaomi's post-training benchmark comparison is published only as an image (assets/benchmark.jpg) on the model card, with no machine-readable table, so this page does not transcribe it. The vendor's stated results are:

• mimo-v2.5-pro "ranks first among open-source models globally on the GDPVal-AA and ClawEval leaderboards."
• ClawEval: "64% Pass^3 using only ~70K tokens per trajectory."
• GraphWalks at 1M tokens: 0.37 BFS / 0.62 Parents.

The base-model table on the model card (5-shot MMLU 89.4, 5-shot GPQA-Diamond 66.7, 8-shot GSM8K 99.6, 1-shot LiveCodeBench v6 39.6) describes MiMo-V2.5-Pro-Base, not the instruct model, and Xiaomi selected the comparison set. Treat all of the above as vendor claims.

Human preference (arena.ai)

arena.ai — the former LMArena — ranks mimo-v2.5-pro #29 on its text leaderboard with an Elo of 1466.2 across 34,468 votes (rank interval #15–#39). It carries no rank on arena.ai's webdev leaderboard. Artificial Analysis and arena.ai use unrelated scales and must not be cross-compared.

Limitations

• Hard agentic benchmarks stay near the floor. Despite the long-horizon marketing, Artificial Analysis measures 8.7% on 𝜏³-Banking, 4.0% on CritPt, and 2.4% on Apex-Agents.
• A 24.5% measured hallucination rate. AA-Omniscience scores the model at 3.6 with accuracy of 22.6%. Ground factual queries with retrieval.
• The default system prompt enforces PRC law. The chat template baked into tokenizer_config.json injects a system prompt whenever the caller supplies none, and it states: "You are a Chinese AI model and must strictly comply with all applicable laws and regulations of the People’s Republic of China. Do not generate, assist with, or facilitate any content that violates Chinese law." It also asserts a fixed identity ("You have 1T parameters"). Passing an explicit system message overrides it.
• Vendor benchmarks are unauditable. The post-training comparison exists only as a JPEG. There is no published table, no per-benchmark methodology, and no independent replication of the "first among open-source models" claim.
• Serving cost is real. ~1.03 TB of FP8 weights. Xiaomi's own SGLang example runs --tp-size 16 --ep-size 16 across multiple nodes. MIT does not make this cheap to host.
• 1M context is not universal. Among OpenRouter providers, DigitalOcean caps the window at 87,040 tokens and DeepInfra charges $1.00 / $3.00 — more than double Xiaomi's first-party rate.
• Open weights, closed data. MIT covers the weights and code. Training data, data mixture, and the RL environments are not released.
• Enterprise headwinds. Omdia's Lian Jye Su, quoted by InfoWorld, notes that "closed frontier models may still win on generic tasks, and the hardest edge cases," and that "Chinese-origin models can trigger concerns in regulated Western organizations."

Pricing & Access

Self-hosting is free under MIT: XiaomiMiMo/MiMo-V2.5-Pro on Hugging Face, mirrored on ModelScope.

Xiaomi publishes the same rate card in two currencies, one per locale. The USD figures are a fixed conversion of the CNY ones at 0.145 USD per yuan — ¥3 × 0.145 = $0.435, ¥6 × 0.145 = $0.87 — so they are not independently set prices.

A premium MiMo-V2.5-Pro-UltraSpeed tier — "Combines FP4 lossless quantization and DFlash parallel decoding," advertised at "1,000 tokens/s peak" — is priced at $0.0108 / $1.305 / $2.61 (¥0.075 / ¥9 / ¥18), and is early-access only. Its FP4 weights are published as XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash.

Third-party hosting via OpenRouter as xiaomi/mimo-v2.5-pro:

Consumer access runs through Xiaomi MiMo Studio and a "MiMo Claw" subscription, advertised as a limited-time ¥14.9/month offer bundling native OpenClaw integration.

Serving: officially supported by both SGLang and vLLM. Xiaomi recommends temperature=1.0, top_p=0.95.

Ecosystem & Tools

• XiaomiMiMo/MiMo-V2.5-Pro on Hugging Face — MIT weights, architecture table, deployment guides. 101,692 downloads in the 30 days to July 8, 2026, and 698 cumulative likes.
• MiMo-V2.5-Pro launch post — the SysY compiler, video editor, and analog-EDA case studies
• Xiaomi MiMo API Platform — first-party API and Token Plan
• vLLM recipe and SGLang cookbook — the two vendor-endorsed serving paths
• OpenRouter xiaomi/mimo-v2.5-pro — five providers, first-party included
• XiaomiMiMo/MiMo-V2.5 — the 310B/15B omnimodal sibling (208,300 downloads/30d); XiaomiMiMo/MiMo-V2.5-ASR is published alongside it, while the MiMo-V2.5-TTS series ships only as an API product — it has no Hugging Face repository
• Harness support — Claude Code, OpenCode, and Kilo, per Xiaomi's launch post

Community & Resources

• MiMo-V2.5-Pro — Xiaomi's launch post, dated April 27, 2026
• Xiaomi MiMo-V2.5 series open-sourced — the MIT License announcement
• MiMo-V2.5-Pro on Artificial Analysis — independent Intelligence Index of 42.2 (v4.1)
• Xiaomi releases MIT-licensed MiMo models for long-running AI agents — InfoWorld, with analyst caveats
• Xiaomi open-sources MiMo-V2.5 Pro under MIT license — The Standard
• arena.ai text leaderboard — #29, Elo 1466.2
• Compare with DeepSeek V4, Kimi K2.6, GLM-5.2, Ling-2.6-1T, MiniMax-M3, and LongCat-2.0