Editorial illustration for Xiaomi launches MiMo‑V2.5‑Pro and V2.5, matching benchmarks at lower token cost
Xiaomi MiMo-V2.5 Slashes AI Token Costs Dramatically
Xiaomi launches MiMo‑V2.5‑Pro and V2.5, matching benchmarks at lower token cost
Xiaomi’s latest AI rollout—MiMo‑V2.5‑Pro and its lighter‑weight sibling MiMo‑V2.5—promises the same headline‑grabbing benchmark scores as leading frontier models while slashing the token cost required for inference. The company frames the achievement as a practical win for developers who need high‑quality code generation without the expense of running massive models. But raw numbers only tell part of the story.
To back its claim, Xiaomi leans on an internal testing framework that mimics the kinds of tasks software engineers actually face, from digging through repositories to stitching together full projects. The suite also stresses the importance of agentic workflows, where code‑writing assistants operate alongside tools like Claude Code. By grounding performance claims in a real‑world developer context, Xiaomi hopes to demonstrate that its new models aren’t just theoretically strong—they’re built for the day‑to‑day pressures of software development.
*MiMo Coding Bench is Xiaomi's in-house evaluation suite designed to assess models on real-world developer tasks within agentic frameworks like Claude Code. It covers repo understanding, project building, code review, structured artifact generation, planning, SWE, and more. V2.5-Pro leads the field o*
MiMo Coding Bench is Xiaomi's in-house evaluation suite designed to assess models on real-world developer tasks within agentic frameworks like Claude Code. It covers repo understanding, project building, code review, structured artifact generation, planning, SWE, and more. V2.5-Pro leads the field on this benchmark, and Xiaomi explicitly positions it as a drop-in backend for scaffolds including Claude Code, OpenCode, and Kilo.
MiMo-V2.5: Native Omnimodal at Half the Cost While V2.5-Pro targets the hardest long-horizon agentic tasks, MiMo-V2.5 is a major step forward in agentic capability and multimodal understanding. With native visual and audio understanding, MiMo-V2.5 reasons seamlessly across modalities, surpasses MiMo-V2-Pro in agentic performance, and supports up to 1 million tokens of context.
Xiaomi's MiMo‑V2.5‑Pro and V2.5 arrive with benchmark scores that sit alongside current frontier models, yet at a fraction of the token price. Results look promising. The company says the models are accessible now via API and priced competitively, which could lower entry barriers for developers.
However, the claim rests on Xiaomi's internal MiMo Coding Bench, an evaluation suite that mirrors real‑world developer tasks within agentic frameworks such as Claude Code. That suite measures repo understanding, project building, code review, structured artifact generation, planning, and software‑engineering workflows, and the press release notes V2.5‑Pro leads the field on these metrics. Still, external verification of those results is pending, and it is unclear whether the lower token cost will hold up under diverse workloads outside the test environment.
Does the lower token price translate to broader use? The models’ agentic focus suggests they can handle multi‑step tasks rather than isolated queries, a shift that may matter for certain applications. Whether the broader AI community will adopt these offerings remains to be determined, given the limited public data beyond Xiaomi's own benchmarks.
Further Reading
- MiMo-V2.5-Pro - Xiaomi
- MiMo-V2.5: Chat & Inference API - LLMBase
- MiMo-V2-Pro: Everything you need to know - Artificial Analysis
Related Reading
Research & Benchmarks Hyperparameter Tuning Reaches 0.9617 Accuracy in 64.59 Seconds
Research & Benchmarks Pharma Cautious as AI Promises Faster Drug Discovery and Smarter Trials
Research & Benchmarks Google AI Advisors Let Users Probe Performance with Conversational “Why” Queries
From Archives Designing Production-Grade CAMEL Multi-Agent Systems: Start with Docs and GitHub
From Archives Multi-agent AI systems incur higher token costs than single agents in practice
Common Questions Answered
How do the MiMo-V2.5-Pro and V2.5 models differ in performance and cost?
The MiMo-V2.5-Pro and V2.5 models achieve similar benchmark scores to leading frontier models while significantly reducing token inference costs. The V2.5-Pro version leads the field on Xiaomi's internal MiMo Coding Bench, offering high-performance code generation capabilities at a more affordable price point for developers.
What is the MiMo Coding Bench and how does it evaluate AI models?
The MiMo Coding Bench is Xiaomi's in-house evaluation suite designed to assess AI models on real-world developer tasks within agentic frameworks. It comprehensively tests models across multiple dimensions including repo understanding, project building, code review, structured artifact generation, planning, and software engineering challenges.
Which development frameworks are compatible with Xiaomi's new AI models?
Xiaomi explicitly positions the MiMo-V2.5-Pro as a drop-in backend for scaffolds including Claude Code, OpenCode, and Kilo. The models are now accessible via API and are designed to provide seamless integration for developers working across different coding environments.