MiMo-V2.5-Pro

2026-04-27 · Publisher : Xiaomi · Domain : Artificial Intelligence

Xiaomi opens up MiMo-V2.5-Pro, a MoE with 1.02 trillion parameters of which 42 billion are active, with a one-million-token context, under the MIT license, aimed at long-running agent and software engineering tasks.

Xiaomi opens up MiMo-V2.5-Pro, a MoE with 1.02 trillion parameters of which 42 billion are active, with a one-million-token context, under the MIT license, aimed at long-running agent and software engineering tasks. ### Context MiMo-V2.5-Pro succeeds MiMo-V2-Pro and extends the architecture of MiMo-V2-Flash (December 2025) to the trillion-parameter scale. In spring 2026, the closed references are Claude Opus 4.6, Gemini 3.1 Pro and GPT-5.4, which Xiaomi uses as its comparison. The model is published at the same time as MiMo-V2.5 (310 billion parameters, 15 billion active, omnimodal). ### What the model brings - One-million-token context: on GraphWalks, Xiaomi reports 0.56 (breadth-first traversal) and 0.92 (parents) at 512,000 tokens, then 0.37 and 0.62 at 1 million, whereas MiMo-V2-Pro collapses to 0 at 1 million. - Hybrid attention carried over from MiMo-V2-Flash with a ratio of 6 sliding-window layers (128 tokens) for one global layer, and three multi-token prediction (MTP) modules. - Post-training combining SFT, large-scale agent reinforcement learning and multi-teacher on-policy distillation (MOPD). - Token efficiency: on Claw-Eval, Xiaomi reports a 64% success rate (pass^3) with about 70,000 tokens per trajectory, i.e. 40 to 60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro and GPT-5.4 at a comparable level (vendor claim). - Supported from launch by SGLang and vLLM, with hardware partnerships (AWS, AMD, T-Head, Enflame) cited by VentureBeat. ### Architecture and specifications - Architecture: MoE with 1.02 trillion parameters in total, 42 billion active; 70 layers (1 dense, 69 MoE), 384 routed experts of which 8 per token; 10 global attention layers and 60 sliding-window layers. - Context: up to 1 million tokens (256,000 for the base model). Precision: mixed FP8 (E4M3). - Training: 27 trillion tokens, native sequences of 32,000 tokens. Modalities: text for the Pro; MiMo-V2.5 is omnimodal. - License: MIT, commercial use, resumed training and fine-tuning without authorization. - Weights: MiMo-V2.5-Pro and MiMo-V2.5-Pro-Base on Hugging Face and ModelScope; knowledge cutoff not disclosed. ### Cost and access Official API pricing reported by VentureBeat: 1 dollar per million input tokens and 3 dollars for output up to 256,000 tokens of context for MiMo-V2.5-Pro (MiMo-V2.5: 0.40 and 2.00 dollars). For self-hosting, the model weighs about 1 terabyte of files on Hugging Face (repository statistic); the minimum hardware configuration is not published in the sources reviewed. | Item | Price (USD per million tokens) | | --- | --- | | Input (context up to 256,000) | 1.00 | | Output (context up to 256,000) | 3.00 | | Cache and tier beyond 256,000 | Not disclosed | ### Benchmarks | Benchmark | MiMo-V2.5-Pro | Claude Opus 4.6 | Gemini 3.1 Pro | GPT-5.4 | | --- | --- | --- | --- | --- | | SWE-Bench Pro | 57.2 | 57.3 | 54.2 | 57.7 | | Terminal-Bench 2.0 | 68.4 | 65.4 | 68.5 | 75.1 | | Claw-Eval (pass^3) | 63.8 | 70.4 | 57.8 | 60.3 | | GDPVal-AA (Elo) | 1581 | 1606 | 1317 | 1674 | | τ3-bench | 72.9 | 72.4 | 67.1 | 72.9 | | Humanity's Last Exam (without tools) | 34.0 | 40.0 | 44.4 | 42.7 | | Humanity's Last Exam (with tools) | 48.0 | 53.0 | 51.4 | 58.7 | | MiMo Coding Bench (internal) | 73.7 | 77.1 | 67.8 | - | Source: chart data from the official MiMo-V2.5-Pro announcement page (mimo.xiaomi.com). Scores published by Xiaomi; no third-party measurement included here. ### Versus closed models On the agent and coding benchmarks published by Xiaomi, MiMo-V2.5-Pro is within two points of Claude Opus 4.6 on SWE-Bench Pro (57.2 versus 57.3) and ahead of Gemini 3.1 Pro on that same test (54.2) and on Claw-Eval (63.8 versus 57.8). It is behind on GDPVal-AA (1581 versus 1606 for Opus 4.6 and 1674 for GPT-5.4), on Terminal-Bench 2.0 against GPT-5.4 (68.4 versus 75.1) and on pure reasoning: 34.0 on Humanity's Last Exam without tools versus 40.0 to 44.4 for the three closed models. On price, VentureBeat reports an output price of 3 dollars per million tokens, and the vendor highlights a lower cost per task thanks to token efficiency (vendor claim). The open choice takes on a concrete meaning here: a trillion-parameter-class model, under MIT, usable in-house, whose gap with the closed references of the same period is on the order of a few points on agent coding and a few tens of Elo points on office work, but more marked on pure reasoning. ### Key takeaways MiMo-V2.5-Pro takes Xiaomi from the 300-billion scale to the trillion-parameter scale, with a million tokens of context and an MIT license. According to the vendor's figures, it sits within a few points of the spring 2026 closed models on agent coding, but remains behind on reasoning without tools.

Tags : LLM, Open Source, Xiaomi, MiMo, MIT, MoE, 1 M de tokens