#21Xiaomi (MiMo)

MiMo-V2.6-Flash

The smaller, production-oriented model in the V2.6 series: a 309B-parameter mixture-of-experts with 15B active parameters. It has the same omnimodal input and 1M-token context as Pro. It is released under MIT on Hugging Face and costs less than Pro on Xiaomi's API. Xiaomi reports 65.68 on DeepSWE v1.1.

• 309B total / 15B active MoE
• 1M-token context
• Text, image, video and audio input
• Lab-stated DeepSWE v1.1 65.68

AccessOpen weights on Hugging Face (MIT; MiMo-V2.6-Flash-RL and -MOPD checkpoints); Xiaomi MiMo API platform; AI Studio; OpenRouter.

Price$0.14 input (cache miss) / $0.28 output per 1M tokens; cached input $0.0028

MiMo-V2.5-Pro

Xiaomi's April 2026 flagship: a 1.02T-parameter mixture-of-experts with 42B active parameters and a 1M-token context, aimed at long-running agentic coding. Xiaomi says it uses 40 to 60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro and GPT-5.4 at comparable capability. The weights were released openly.

• 1.02T total / 42B active MoE
• 1M-token context
• Lab-stated ClawEval 64% Pass^3
• Open weights on Hugging Face

AccessOpen weights on Hugging Face (Xiaomi says 'permissive license'; MIT per Wikipedia); Xiaomi MiMo API platform (billed at 2x the MiMo-V2.5 rate); AI Studio.

MiMo-V2.5

A 310B-parameter mixture-of-experts model with 15B active parameters, native vision and audio understanding, and a 1M-token context. It was released openly alongside V2.5-Pro. Xiaomi says it matches V2.5-Pro on internal coding benchmarks at lower cost and is competitive with Gemini 3 Pro on video tasks.

• 310B total / 15B active MoE
• 1M-token context; 128K max output
• Text, image, video and audio understanding
• Lab-stated Claw-Eval 62.3 (general subset)

AccessOpen weights on Hugging Face (MIT per Wikipedia); Xiaomi MiMo API platform; AI Studio.

Price$0.14 input (cache miss) / $0.28 output per 1M tokens; cached input $0.0028

MiMo-V2.5-TTS series (TTS, VoiceDesign, VoiceClone)

A text-to-speech series in three variants. The base model offers stock voices that follow style instructions, VoiceDesign creates new voices from text descriptions, and VoiceClone reproduces a voice from an audio sample. It supports Chinese and English, with inline emotion tags and screenplay-style input for control over delivery.

• Three variants: TTS, VoiceDesign, VoiceClone
• Chinese and English
• Natural-language style direction and inline audio tags
• Free via API for a limited time

AccessXiaomi MiMo API platform and MiMo Studio (free for a limited time). Licence not stated; not listed as open weights.

MiMo-V2.5-ASR

An open-source speech-recognition model for Chinese and English that also handles eight Chinese dialects, code-switching, song lyrics, noisy audio and multiple speakers, with native punctuation. Xiaomi reports a word error rate of 2.52 on AiShell-2 and a 5.73 average on the Open ASR Leaderboard for English.

• Chinese and English plus 8 Chinese dialects
• Handles code-switching, lyrics, multi-speaker audio
• Lab-stated AiShell-2 WER 2.52
• Open weights on Hugging Face

AccessOpen weights on Hugging Face (XiaomiMiMo/MiMo-V2.5-ASR); code on GitHub.

MiMo-V2-Pro

Xiaomi's first trillion-parameter model: a mixture-of-experts with 42B active parameters and up to a 1M-token context, offered only through Xiaomi's API. It launched with MiMo-V2-Omni and MiMo-V2-TTS. Xiaomi said it ranked eighth globally on the Artificial Analysis Intelligence Index. The open MiMo-V2.5-Pro replaced it a month later.

• >1T total / 42B active parameters
• Up to 1M-token context
• Lab-stated ClawEval 61.5, PinchBench 81.0
• API-only, no open weights

AccessXiaomi MiMo API platform only; not released as open weights.

Price$1 input / $3 output per 1M tokens up to 256K context; $2 / $6 for 256K-1M

MiMo-V2-Flash

Xiaomi's first large open-weights model: a 309B-parameter mixture-of-experts with 15B active parameters and a 256k context. It is built for reasoning, coding and agent tasks and Xiaomi cites about 150 tokens per second. At release Xiaomi reported 73.4% on SWE-Bench Verified, the top open-source score.

• 309B total / 15B active MoE
• 256k-token context
• Lab-stated SWE-Bench Verified 73.4%
• MIT-licensed open weights

AccessOpen weights on Hugging Face (MIT); Xiaomi MiMo API platform; AI Studio.

Price$0.10 input / $0.30 output per 1M tokens (at launch)