MiMo-V2.5

MiMo-V2.5 is presented on its official Hugging Face model card as a native omnimodal system that handles text, image, video, and audio understanding within a single architecture, extending a MiMo-V2-Flash backbone with dedicated vision and audio encoders. The architecture inherits a hybrid attention design that interleaves sliding window attention and global attention in a 5:1 ratio with a 128-token sliding window, a structural choice that the documentation links to roughly a six-fold reduction in KV-cache storage while still supporting long-context reasoning. The card positions the model around agentic workflows and multimodal perception, treating unified perception and tool-driven reasoning as core use cases rather than narrow text tasks.

The repository also carries an explicit configuration update notice: users who pulled the model before a specific commit are asked to re-download or manually refresh the configuration and tokenizer files, since outdated configuration can lead to degraded behavior, which is worth noting for anyone reproducing results. Community activity on the NVIDIA developer forums reflects this deployment interest, including a thread devoted to an FP4-quantized variant targeted at DGX Spark and GB10 hardware. Taken together, the model fits practitioners who want an open-weight omnimodal base for long-context agentic applications and who are comfortable tracking upstream configuration changes to keep inference aligned with the latest release.

Hugging FaceXiaomiMiMo/MiMo-V2.5mimo

Open provider page Provider docs

Quick Info

Provider: Hugging Face
Model key: XiaomiMiMo/MiMo-V2.5
Release date: Apr 22, 2026
Last updated: Apr 22, 2026
Knowledge cutoff: 2024-12
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.40
Output token cost: $2.00

Limits

Output tokens: 131,072 tokens
Context window: 262,144 tokens

Transparent token rates

Compare MiMo-V2.5 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about MiMo-V2.5

Deep Infra

CoverageBenchmark

Xiaomi's MiMo-V2.5 is described as a native omnimodal model delivering Pro-level agentic performance at roughly half the inference cost, and surpassing MiMo-V2-Omni in multimodal perception across image and video tasks. Kilo Code lists a 1,048,576-token context window, 131,072-token max output, $0.14 per 1M input via OpenRouter routing, and an Oct 21 retirement date. On coding benchmarks, MiMo-V2.5 scores 56.8% on the AA Coding Index, 43.9% on SciCode, 41.7% on TerminalBench Hard, 73.0% on long-context reasoning, and 67.1% on instruction-following accuracy. Kilo Code's PinchBench agent evaluation records an 89.7% average across multi-step OpenClaw tasks, with strong category scores in coding at 97.8% and productivity at 96.2%.

Deep Infra

CoverageBenchmark

Xiaomi's MiMo-V2.5 is a 310B-parameter open-weights model with 15B active parameters per token, supporting text and image input with text output, a 1M-token context window, an MIT license, and a reasoning-capable architecture. Artificial Analysis scores it 25 on the Intelligence Index, placing it above the comparable-model median of 18 for its open-weights size class. The page also flags MiMo-V2.5 as deprecated and superseded by Xiaomi's MiMo-V2.6-Pro release. MiMo-V2.5 generates 56.1 output tokens per second, pricing at $0.14 per 1M input tokens and $0.28 per 1M output tokens, with a reported 98% cache discount on Artificial Analysis. The model was released in April 2026, and Artificial Analysis recommends considering the newer MiMo-V2.6-Pro instead for forward-looking workloads.

Deep Infra

Coverage

Xiaomi released MiMo-V2.5, an omnimodal model extending the MiMo-V2-Flash backbone with dedicated vision and audio encoders. It natively handles text, image, video, and audio in a unified architecture, targeting strong agentic workflows and multimodal perception. The sparse Mixture-of-Experts model has 310B total parameters with 15B activated and supports a 1M-token context window. It uses a hybrid attention design with a 5:1 SWA/GA ratio, a 729M-param ViT, a 261M audio transformer, 329M MTP modules, and FP8 pre-training on 48T tokens.