Xiaomi MiMo-V2-Flash: 309B Mixture-of-Experts With Multi-Token Prediction

What MiMo-V2-Flash is

MiMo-V2-Flash is Xiaomi's speed-oriented mixture-of-experts model: 309 billion total parameters with 15 billion active per token, MIT-licensed and self-hostable, with a 256K token context window. Xiaomi describe it as built for high-speed reasoning and agentic workflows, using a hybrid attention architecture together with Multi-Token Prediction to cut inference cost — MTP being the technique of predicting several tokens per forward pass rather than one. At roughly $0.10 per million input tokens and $0.30 output through third-party providers, it is among the cheapest models on this site.

Training note

Xiaomi report that large-scale reinforcement learning on code agents generalised to other domains, improving performance on mathematics and general agent tasks as well. They report it as an experimental observation from their own work.

LMArena

On LMArena's text leaderboard MiMo V2 Flash has 1386 Elo (rank 187, thinking effort); by category, 1432 Elo in Coding (rank 186, thinking effort), 1372 Elo in Math (rank 200, thinking effort), 1415 Elo in Expert (rank 168, thinking effort) and 1336 Elo in Creative Writing (rank 199, thinking effort).

MiMo-V2-Flash and MiMo-V2.5

MiMo-V2-Flash has 309 billion total parameters and activates 15 billion per token, with a 256K window. MiMo-V2.5 has a 1M window and a mostly sliding-window attention stack. Both are MIT-licensed.

Architecture

MiMo-V2-Flash stacks eight hybrid blocks, each of five sliding-window attention layers followed by one global attention layer, with a 128-token window; Xiaomi state this reduces KV-cache storage by nearly six times, using a learnable attention sink bias to hold long-context performance. The multi-token prediction module uses dense feed-forward layers and sliding-window attention at 0.33 billion parameters per block, and Xiaomi say it triples generation speed through self-speculative decoding. Pre-training covered 27 trillion tokens in FP8 mixed precision at a native 32K sequence length, with the context supporting 256K. Post-training used Multi-Teacher On-Policy Distillation, in which domain-specific teacher models supervise the student at every token on its own generated responses, followed by large-scale agentic reinforcement learning.

Xiaomi’s figures

For the post-trained model Xiaomi report SWE-Bench Verified 73.4, SWE-Bench Multilingual 71.7, Terminal-Bench 2.0 38.5, AIME 2025 94.1, GPQA Diamond 83.7, LiveCodeBench v6 80.6, LongBench V2 60.6 and τ²-Bench 80.3, with BrowseComp 45.4 rising to 58.3 with context management. The base model holds 96.7 on a multi-needle retrieval test at 256K, and in Xiaomi's base-model table records MMLU-Pro 73.2, GPQA Diamond 55.1, MATH 71.0, BigCodeBench 70.1 and LiveCodeBench v6 30.8, against Kimi-K2 Base's 69.2, 48.1, 70.2, 61.7 and 26.3. On writing, the post-trained model records 86.2 on Arena-Hard creative writing and 54.1 on Arena-Hard hard prompts. A technical report is published in Xiaomi's GitHub repository. Xiaomi describe Multi-Teacher On-Policy Distillation as formulating knowledge distillation as a reinforcement-learning process, with rewards derived from the divergence between student and teacher distributions, which they state makes the process resistant to reward hacking. The model is MIT-licensed. Xiaomi publish MiMo-V2-Flash-Base and MiMo-V2-Flash on Hugging Face, open-source the three-layer MTP weights, and offer the model through the Xiaomi MiMo API Platform and MiMo Studio.

Official Resources