Xiaomi MiMo-V2.5 Pro Goes Open Source: 1T-Parameter Agent Model Beats DeepSeek-V4

Key Takeaways

Xiaomi's AI research division MiMo released its most advanced model to date on April 28, 2026, open-sourcing MiMo-V2.5 and MiMo-V2.5-Pro under the MIT License. The release marks a significant step forward from the earlier MiMo-V2-Pro, with the Pro variant now leading the open-source field on agentic coding benchmarks while consuming substantially fewer tokens than competing frontier models. Alongside the model release, Xiaomi launched the MiMo Orbit developer program offering 100 trillion free usage tokens to selected applicants over 30 days.

Feature Overview

1. Architecture and Scale

MiMo-V2.5-Pro is built on a Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion active parameters per forward pass. The model operates with a 1-million-token context window and incorporates a lightweight multi-token prediction module that triples output throughput compared to standard autoregressive generation. Training used mixed-precision FP8 across 27 trillion tokens.

The standard MiMo-V2.5 is a native omnimodal model supporting text, images, video, and audio, with 310 billion total parameters and 15 billion active during inference.

2. Benchmark Performance

On ClawEval, the primary benchmark for agentic coding tasks, MiMo-V2.5-Pro achieves a 63.8% Pass@3 rate using approximately 70,000 tokens per trajectory. This efficiency advantage is significant: comparable results from Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 require 40 to 60 percent more tokens per trajectory. On the GDPVal-AA benchmark for economically valuable knowledge work, the model ranks first among open-source models and outperforms DeepSeek-V4-Pro.

Practical demonstration results reported by Xiaomi include a complete SysY-to-RISC-V compiler written in Rust over 672 tool calls in 4.3 hours, passing 233 of 233 hidden test cases. In a second demonstration, the model produced an 8,192-line desktop video editor over 1,868 tool calls spanning 11.5 hours of autonomous operation.

3. MIT License and Commercial Access

The MIT License chosen for MiMo-V2.5 Pro imposes no restrictions on commercial deployment, continued training on proprietary data, or distribution. Developers can fine-tune on private datasets, integrate the model into commercial products, and build derivative models without seeking additional authorization or paying revenue-based fees. This is a materially more permissive license than the custom community licenses used by many competing open-weight models.

4. MiMo Orbit Developer Program

Simultaneous with the model release, Xiaomi opened applications for the MiMo Orbit program. Selected applicants receive access to 100 trillion free inference tokens over 30 days through the MiMo API. Xiaomi has not disclosed the selection criteria or the total number of participants accepted. The program is positioned as a mechanism for developers to evaluate the model at scale before committing to production deployment.

5. Relationship to MiMo-V2-Pro

The V2.5 series represents a substantial improvement over MiMo-V2-Pro, which was revealed in March 2026 as the anonymous model that had topped OpenRouter usage charts under the codename Hunter Alpha. Where V2-Pro scored 61.5 on ClawEval, V2.5-Pro reaches 63.8 at meaningfully lower token cost. Xiaomi states that V2.5 delivers improvements across general agentic capabilities, complex software engineering, and long-horizon task execution.

Usability Analysis

For developers building autonomous coding agents or long-running task pipelines, MiMo-V2.5-Pro offers a compelling combination of high task completion rates and low token consumption. The MIT license removes the legal friction that accompanies models with custom terms, making it straightforward to incorporate into commercial workflows.

Access is available through the MiMo API and Hugging Face model hub. Weights can be downloaded for local deployment, though running a 1-trillion-parameter MoE model at full scale requires substantial GPU infrastructure. Developers working with smaller hardware budgets may find the standard MiMo-V2.5 omnimodal variant more practical for local inference.

The MiMo Orbit program's free token allocation provides a low-risk path to evaluate V2.5-Pro at production token volumes before committing to paid API usage.

Pros

1. MIT license with no commercial restrictions enables immediate production deployment and fine-tuning without legal overhead
2. Leading ClawEval agentic performance at 63.8% Pass@3 outperforms DeepSeek-V4-Pro among open-source models
3. Token efficiency advantage of 40-60% versus Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable capability levels reduces operating costs directly
4. 1-million-token context window supports processing of large codebases and extended autonomous sessions
5. 100T free tokens through MiMo Orbit lowers evaluation friction for new adopters

Limitations

1. Infrastructure requirements for running the 1T-parameter model locally are substantial, limiting full-scale self-hosting to well-resourced teams
2. Limited enterprise support ecosystem compared to established Western providers; documentation and SLA commitments remain less mature
3. Geopolitical considerations around Chinese AI models may restrict adoption in government-regulated industries or organizations with relevant compliance requirements
4. MiMo Orbit selection criteria are not disclosed, so guaranteed access to the free token program is uncertain for all applicants

Outlook

MiMo-V2.5-Pro's open-source release continues a pattern of Chinese AI labs delivering frontier-adjacent performance at aggressive cost and access terms. With DeepSeek-V4-Pro and now MiMo-V2.5-Pro both available under permissive licenses, the effective cost of running capable open-weight agents has dropped sharply in 2026.

Xiaomi's willingness to open-source a model at this scale under MIT terms suggests the company is prioritizing developer ecosystem growth over near-term API revenue, likely with the goal of establishing MiMo as the default infrastructure for Xiaomi's own device and IoT AI applications. If Xiaomi follows the MiMo Orbit program with sustained documentation investment and stable API availability, it has a credible path to becoming a significant player in the developer AI stack outside China.

The token efficiency story is particularly important for the agentic computing segment, where trajectory length directly determines operating cost. A 40-60% reduction in tokens-per-task is not a marginal improvement; at scale, it determines which models are economically viable for high-volume agent deployments.

Conclusion

MiMo-V2.5-Pro is a technically strong open-source release that leads the field on agentic coding benchmarks while delivering meaningful token efficiency advantages over proprietary frontier models. The MIT license, combined with the MiMo Orbit free token program, gives developers a practical path to evaluate and deploy the model without commercial friction. For teams building autonomous coding agents and long-horizon task pipelines, MiMo-V2.5-Pro is a serious evaluation candidate that justifies direct testing against proprietary alternatives.

Editor's Verdict

Xiaomi MiMo-V2.5 Pro Goes Open Source: 1T-Parameter Agent Model Beats DeepSeek-V4 earns a solid recommendation within the other llm space.

The strongest case for paying attention is MIT license with no commercial restrictions enables immediate production use and custom fine-tuning, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, leads open-source field on ClawEval agentic coding benchmark at 63.8% Pass@3 adds practical value rather than just headline appeal. The broader signal worth registering is straightforward: MIT license removes all commercial and fine-tuning restrictions, making MiMo-V2.5-Pro immediately viable for production deployment without legal overhead. On the other side of the ledger, full 1T-parameter local deployment requires significant GPU infrastructure beyond most individual developers is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, enterprise support ecosystem, documentation depth, and SLA commitments lag behind established Western providers narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.