Xiaomi has introduced its MiMo-V2.5 model family, adding multimodal capabilities and advancing its push into top-tier AI systems.
Summary
- Xiaomi has launched the MiMo V2.5 and V2.5 Pro models, combining text, image, audio, and video capabilities into a single system.
- MiMo V2.5 Pro has delivered near top-tier benchmark results, resolving 57.2% of tasks on the SWE bench Pro while competing with leading AI models.
- The company has priced the models lower while improving token efficiency, using up to 42% fewer tokens than comparable systems for similar performance.
According to Xiaomi, the new MiMo-V2.5 and MiMo-V2.5-Pro models combine image, audio, and video processing into a single system, bringing features that were previously split across separate models into one integrated release.
What sets MiMo-V2.5 apart?
The earlier MiMo-V2-Pro focused on text and code, while multimodal functions were handled by a different, lower-performing model. That separation no longer exists. MiMo-V2.5 merges those capabilities, allowing users to process images, videos, and audio without switching tools or models.
Users can upload a photo and request suggestions, analyse video tutorials for step-by-step guidance, or extract action points from recorded meetings, all within the same system.
Xiaomi describes the Pro version as “a major leap from MiMo-V2-Pro in general agentic capabilities, complex software engineering, and long-horizon tasks,” adding that it now matches leading systems such as Claude Opus 4.6 and GPT-5.4 across most coding and agent benchmarks.
Performance, pricing, and positioning
MiMo-V2.5-Pro is built for demanding workflows. Xiaomi says it can “autonomously complete professional tasks involving 1,000+ tool calls, work that would take human experts days.”
Running at 60 to 80 tokens per second, the Pro model is priced at $1.00 per million input tokens and $3.00 per million output tokens. The base MiMo-V2.5 model targets everyday use, offering faster speeds of 100 to 150 tokens per second at lower costs of $0.40 input and $2.00 output. Both models support a 1M-token context window, allowing them to handle large datasets or extended conversations.
Benchmark data places the Pro model near the top tier. On SWE-bench Pro, it resolves 57.2% of tasks, more than double the typical average of around 25%. Results on τ3-bench and ClawEval place it close to leading models, though performance drops on more complex reasoning tests such as Humanity’s Last Exam, where it scores 48.0% compared to GPT-5.4’s 58.7%.
Efficiency has emerged as a key differentiator. Xiaomi says MiMo-V2.5-Pro uses 42% fewer tokens than Kimi K2.6 for similar results, while the base model consumes nearly half the tokens required by comparable systems. For developers operating at scale, lower token usage directly reduces costs.
Rapid rollout and ecosystem push
Recent releases show a steady cadence. Xiaomi introduced MiMo-V2-Flash in late 2025, followed by the V2-Pro, Omni, and TTS models in March, before unveiling the V2.5 series.
Lei Jun announced an $8.7B investment in AI over the next three years, and activity since then suggests that deployment has already accelerated.
Platform data adds context. Xiaomi’s models accounted for about 21% of OpenRouter traffic as of early April, with usage rising more than 42% in a single week. That growth followed a period of free access through the Hermes agentic AI tool, which expanded visibility and adoption.
Pricing updates accompany the launch. Xiaomi removed additional charges for using the full 1M-token context window and reset user credits as part of the rollout. The models are available through the MiMo API, while access through AI Studio remains limited.
Xiaomi said future models will focus on “deeper reasoning, tighter tool integration, and richer real-world grounding,” suggesting that another release may arrive sooner than expected.
