Xiaomi releases MIT‑licensed MiMo models for long‑running AI agents

MiMo‑V2.5 targets developers building autonomous coding

With a 1‑million‑token context window and sparse MoE design, MiMo‑V2.5 targets developers building autonomous coding and workflow agents.

Xiaomi has released and open-sourced MiMo-V2.5 and MiMo-V2.5-Pro under the MIT License, giving developers another potentially lower-cost option for building AI agents that can run longer tasks such as coding and workflow automation.

Both models support a 1-million-token context window, the company said. MiMo-V2.5-Pro is designed for complex agent and coding tasks, while MiMo-V2.5 is a native omnimodal model that supports text, images, video, and audio.

The release comes as agentic AI workloads are putting new pressure on enterprise AI budgets. These systems can burn through large numbers of tokens as they plan, call tools, write code, and recover from errors, making cost and deployment control increasingly important for developers.

By using the MIT License, Xiaomi said it is allowing commercial deployment, continued training, and fine-tuning without additional authorization. Tulika Sheel, senior vice president at Kadence International, said the MIT License can make it attractive. “It allows enterprises to freely modify, deploy, and commercialize the model without restrictions, which is rare in today’s AI landscape,” Sheel said.

“On ClawEval, V2.5-Pro lands at 64% Pass^3 using only ~70K tokens per trajectory — roughly 40–60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable capability levels,” Xiaomi said in a blog post.

The models use a sparse mixture-of-experts (MoE) design to manage compute costs. The 310-billion-parameter MiMo-V2.5 activates only 15 billion parameters per request, while the 1.02-trillion-parameter Pro version activates 42 billion. Xiaomi said the Pro model’s hybrid attention design can reduce KV-cache storage by nearly seven times during long-context tasks.

Xiaomi cited several long-horizon tests, including a SysY compiler in Rust that MiMo-V2.5-Pro completed in 4.3 hours across 672 tool calls, passing 233 of 233 hidden tests. It also said the model produced an 8,192-line desktop video editor over 1,868 tool calls across 11.5 hours of autonomous work.

Will enterprises adopt MiMo?

Whether Xiaomi’s MiMo-V2.5 models can gain adoption among enterprise developers over closed frontier models for agentic coding and automation workloads will depend on how enterprises evaluate performance, cost, and risk.

“When assessing Xiaomi’s MiMo-V2.5 and its variants, enterprise developers should look at the total cost of ownership,” said Lian Jye Su, chief analyst at Omdia. “The TCO consists of token efficiency, cost per successful task, and the absence of licensing costs associated with proprietary models. Closed frontier models may still win on generic tasks, and the hardest edge cases, but open-weight models excel in agentic work that is high-volume in nature.”

Pareekh Jain, CEO of Pareekh Consulting, said enterprises should assess MiMo-V2.5 less as a replacement for Claude or GPT and more as a cost-efficient agent model for high-token workloads.

“The key benchmark signal is not just accuracy, but tokens per successful task,” Jain said. “Frontier models often reach higher success rates on complex coding benchmarks, but do so with massive reasoning overhead. MiMo-V2.5 is designed for Token Efficiency, meaning it achieves comparable results with significantly fewer input and output tokens.”

Jain said that could make MiMo-like models useful as “economic workhorses” for repetitive coding, QA, migration, documentation, testing, and automation workloads, while closed frontier models remain the quality ceiling for the hardest tasks.

Ashish Banerjee, senior principal analyst at Gartner, said models like MiMo could materially shift enterprise AI economics for long-horizon agents.

“When tasks stretch into millions of tokens, metered proprietary APIs stop looking like a convenience and start looking like a tax on iteration,” Banerjee said. “By contrast, MiMo’s MIT license, open weights, 1M-token context window, and relatively low pricing make private-cloud or self-hosted deployment strategically credible.”

However, Banerjee said this does not mean enterprises will abandon proprietary APIs.

“Enterprises will continue to use proprietary APIs for frontier accuracy and low-operations consumption, while shifting scaled, repeatable agent workflows toward open models where cost predictability, data control, and customization matter more,” Banerjee said. “In short, long-horizon, high-volume agentic AI will evolve into a hybrid market, with open models like MiMo breaking pure API dependence.”

Su added that adoption may face challenges because Chinese-origin models can trigger concerns in regulated Western organizations.

Credit: Prasanth Aby Thomas, Computer World