Xiaomi MiMo

Xiaomi's entry into open reasoning models: MiMo-7B, released across base, SFT, RL-zero and RL checkpoints with a technical report. The RL model reports wins over o1-mini and QwQ-32B-Preview on AIME and LiveCodeBench v5, with pass@k scores above a 32B baseline.

Xiaomi and TileRT report 1,000+ tokens per second on a 1T-parameter MoE model using a single standard 8-GPU node, via FP4 quantization and DFlash block-masked parallel speculative decoding with a compiler tuned to both. Weights for the FP4 plus DFlash checkpoint are open.

Xiaomi's MiMo team put MiMo Claw into general availability: an agent product on its flagship model with Kingsoft Office integration, pitched at reliable long-running tasks and 40 to 60 percent lower token consumption than comparable solutions.

Xiaomi released MiMo-VL-7B in SFT and RL variants, with GUI grounding strong enough to pass UI-TARS on OSWorld-G at 56.1 and OlympiadBench at 59.4. The recipe: 2.4T tokens of staged image, video and GUI-trajectory pretraining, then mixed on-policy RL. Weights and code are public.

Xiaomi's HySparse interleaves each full attention layer with sparse layers that inherit both token selection and the KV cache from it. At 80B MoE scale only 5 of 49 layers need full attention, cutting KV cache storage roughly 10x while beating the full-attention baseline.

Xiaomi's own run logs for MiMo-V2.5-Pro on long-horizon tasks: a Rust SysY compiler passing 233 of 233 hidden tests in 4.3 hours and 672 tool calls, an 8,192-line video editor over 1,868 calls, and an analog LDO design meeting six specs via an ngspice loop driven through Claude Code.

Xiaomi released MiMo-Audio with the claim that scaling pretraining past 100M hours produces few-shot generalization across audio tasks. The 7B instruct model is reported at the top of MMSU, MMAU, MMAR and MMAU-Pro. Tokenizer, weights and evaluation suite are all open.

Xiaomi shipped MiMo-V2-Pro, the omni-modal MiMo-V2-Omni and a TTS model as a single release, with free trial access routed through OpenClaw, OpenCode, Kilo Code, Cline and BlackBox, later extended by a further week.

Xiaomi's MiMo team published a breakdown of the serving stack behind its V2.5 models, centered on pushing hybrid sliding-window attention as far as it goes: cache pressure, prefill latency and decode throughput for long-context and multimodal requests.

An in-place upgrade to MiMo-VL at the same size: MMMU 70.6, VideoMME 70.8, and a no_think flag that switches reasoning traces off for direct answers. Xiaomi reports its VLM Arena rating rising from 1093.9 to 1131.2. RL and SFT weights are both posted.

Thinking mode support arrived for MiMo-V2-Flash inside Claude Code, Kilo Code and Cline, alongside stability work across scenarios. Free API access ran through January 20, 2026, making a side-by-side against your usual coding model cheap to run.

Two models: MiMo-V2.5-Pro for long-horizon agent and software work, reported near frontier on SWE-bench Pro, Claw-Eval and tau3-Bench, and MiMo-V2.5, a native omni-modal model with a 1M context at roughly half the cost. Both are pitched on tokens spent per unit of score, not score alone.

Xiaomi's MiMo-V2.5 is now available as a backend in the OpenCode coding agent, free for a limited window, with a 1M-token context, reasoning, and image input. A no-cost option for testing a large-context model against your own repository.

Xiaomi released MiMo-V2.6 Pro and Flash, omnimodal models trained via scaled RL. Xiaomi claims Pro matches Claude Opus 5 and GPT-5.6 on agent benchmarks and leads open models on the Artificial Analysis Intelligence Index, with weights and RL environments open-sourced.

Xiaomi cut MiMo-V2.5 API pricing by up to 99% and replaced per-context-length tiers with a single flat rate, effective May 26. Token plans now carry 5x to 8x more usable tokens, existing plan credits are reset, TTS stays free for now, and the 100T builder grant has ended.

Xiaomi's 309B total, 15B active MoE release comes with real architectural disclosure: hybrid 5:1 SWA plus global attention at a 128-token window, attention sink bias as the load-bearing detail, MTP heads reused as speculative draft models, and multi-teacher on-policy distillation.

Xiaomi opened account top-ups on its MiMo API platform and published rates for MiMo-V2-Flash: $0.10 per million input tokens, $0.01 cached, $0.30 output. Billing was not yet enabled, so calls stayed free and balances untouched until a later announced date.

Xiaomi's voice release splits into three TTS models (built-in voices, voice design from a text description, few-sample cloning) sharing inline audio tags for emotion and pacing control, plus an open-sourced MiMo-V2.5-ASR covering dialects, code-switching, noisy and overlapping-speaker audio.

Xiaomi's Token Plan bundles MiMo-V2-Pro, Omni and TTS under one subscription with no five-hour usage windows, separate OpenAI and Anthropic compatible endpoints, regional hosts for China, Europe and elsewhere, and a tp- prefixed key format.

Xiaomi published MiMo-V2.5-Pro and MiMo-V2.5 weights under MIT, allowing commercial deployment, continued training and fine-tuning without further permission. Both carry 1M context. SGLang and vLLM shipped support on release day, alongside adaptation for AMD and several other accelerators.

Xiaomi open-sourced MiMo Code V0.1 under MIT: a terminal coding agent that loads existing Claude Code skills, MCP servers and commands, adds a spec-then-build Compose mode, session self-review, and voice input via MiMo-V2.5-ASR, with multi-provider support beyond its bundled model.

A first-party integration warning: MiMo's thinking mode requires reasoning_content to be replayed on any assistant message containing tool calls, or the API returns 400. Dropping it truncates context and weakens instruction following. Cursor, Codex, Zed and AutoGen were among those affected.