MiMo-V2-Flash

2025-12-16 · Publisher : Xiaomi · Domain : Artificial Intelligence

Xiaomi releases MiMo-V2-Flash, an open-weights model (MIT) with 309 billion parameters, 15 of them active, designed for agents and code, with a hybrid attention that reduces the cache by about six times.

Xiaomi releases MiMo-V2-Flash, an open-weights model (MIT) with 309 billion parameters, 15 of them active, designed for agents and code, with a hybrid attention that reduces the cache by about six times. ### Context At the end of 2025, the race for open models is played out on coding agents and long tasks: Kimi K2 Thinking and DeepSeek-V3.2 set the reference on the open side, Gemini 3 Pro, Claude Sonnet 4.5 and GPT-5 on the closed side. Xiaomi had opened a 7-billion-parameter model in April 2025; MiMo-V2-Flash changes scale and targets the speed/cost ratio rather than raw size. ### What the model brings - Hybrid attention: five sliding-window layers (128 tokens) for one global attention layer, with a learnable attention bias; Xiaomi announces a key-value cache reduced by about six times while preserving a long context. - Native multi-token prediction (MTP): a lightweight module of 0.33 billion parameters per block, three layers of which are published; the vendor states a generation speed multiplied by three. - Post-training by multi-teacher on-policy distillation (MOPD): domain-specific expert models guide the student at each token, which replaces the sparse reward with a dense signal. - Large-scale agent reinforcement learning: more than 100,000 verifiable tasks drawn from real GitHub issues, a Kubernetes cluster of more than 10,000 simultaneous pods and a multimodal verifier for web development. - Base weights and post-trained weights published, with the inference code. ### Architecture and specifications - Architecture: MoE, 309 billion parameters in total, 15 billion active; 8 hybrid blocks of 5 sliding-window layers and 1 global layer. - Training: 27 trillion tokens, FP8 mixed precision, native sequences of 32,000 tokens. - Context: up to 256,000 tokens. Modalities: text. - License: MIT (free commercial use). Weights: MiMo-V2-Flash and MiMo-V2-Flash-Base on Hugging Face. - Knowledge cutoff: not disclosed. ### Cost and access Xiaomi's official API pricing at launch: 0.1 dollars per million input tokens and 0.3 dollars for output. For self-hosting, the vendor recommends SGLang in FP8 precision; the model card's example launches the server with tp-size 8 and dp-size 2, without specifying the GPU model. | Item | Price (USD per million tokens) | | --- | --- | | Input | 0.1 | | Output | 0.3 | | Cache and batch | Not disclosed | ### Benchmarks | Benchmark | MiMo-V2-Flash | Gemini 3.0 Pro | Claude Sonnet 4.5 | GPT-5 High | | --- | --- | --- | --- | --- | | MMLU-Pro | 84.9 | 90.1 | 88.2 | 87.5 | | GPQA-Diamond | 83.7 | 91.9 | 83.4 | 85.7 | | Humanity's Last Exam (without tools) | 22.1 | 37.5 | 13.7 | 26.3 | | AIME 2025 | 94.1 | 95.0 | 87.0 | 94.6 | | LiveCodeBench v6 | 80.6 | 90.7 | 64.0 | 84.5 | | SWE-Bench Verified | 73.4 | 76.2 | 77.2 | 74.9 | | SWE-Bench Multilingual | 71.7 | - | 68.0 | 55.3 | | Terminal-Bench 2.0 | 38.5 | 54.2 | 42.8 | 35.2 | | BrowseComp | 45.4 | - | 24.1 | 54.9 | | τ²-Bench | 80.3 | 85.4 | 84.7 | 80.2 | Source: MiMo-V2-Flash Hugging Face model card, post-training table. Scores published and measured by Xiaomi; the same table includes Kimi K2 Thinking and DeepSeek-V3.2 Thinking (open). ### Versus closed models On agent coding, MiMo-V2-Flash sits within the range of the closed models cited: 73.4 on SWE-Bench Verified against 74.9 for GPT-5 High and 77.2 for Claude Sonnet 4.5, and 71.7 on SWE-Bench Multilingual, ahead of Claude Sonnet 4.5 (68.0) and GPT-5 High (55.3). It remains clearly behind Gemini 3.0 Pro on knowledge (GPQA-Diamond 83.7 against 91.9), competitive coding (LiveCodeBench 80.6 against 90.7) and Terminal-Bench 2.0 (38.5 against 54.2). Web browsing is a weak point: 45.4 on BrowseComp against 54.9 for GPT-5 High. On price, Xiaomi states 0.1 and 0.3 dollars per million tokens; the vendor writes that this cost represents about 3.5% of that of Claude Sonnet 4.5 (vendor claim, not verified by a third party). For the open versus closed choice, the model offers the sovereignty of an internal deployment under an MIT license with a level close to mid-range closed models on agent coding, at the price of lagging on advanced knowledge tasks. ### Key takeaways MiMo-V2-Flash puts a MoE with 15 billion active parameters at the level of mid-range closed models on SWE-Bench, at a very low price and under an MIT license. Its hybrid sliding-window attention and its MTP are the innovations to watch for reducing the cost of long contexts.

Tags : LLM, Open Source, Xiaomi, MiMo, MIT, MoE, Agents