In the open-source LLM camp, a 1T-parameter model under MIT license, purpose-built for Agent workloads, means there’s yet another option developers can drop straight into production.
On April 28, Xiaomi officially open sourced the MiMo-V2.5 series with two models: the Pro version at 1.02T total parameters (42B active) and the standard version at 310B total (15B active). Weights, tokenizer, and model cards are all on Hugging Face under the MIT license — commercial use, continued training, and fine-tuning are all permitted without additional authorization.
Focus: Agent Capability, Not General Chat
MiMo-V2.5’s strategy is clear: rather than competing with GPT and Claude on general-purpose chat across all dimensions, it focuses on Agent scenarios — code generation, tool use, multi-step reasoning.
MiMo-V2.5-Pro on Agent benchmarks:
- SWE-bench Pro: 57.2, approaching Claude Opus 4.6
- Claw-Eval: 63.8
- τ3-Bench: 72.9
The MiMo-V2.5 (310B) scores 49 on the Artificial Analysis Intelligence Index, comparable to GPT-5.5 mini x-high and Grok 4.2. The Pro version reaches 54, on par with GPT-5.3 Codex and Kimi K2.6.
Notably, when completing identical Agent tasks, the Pro version consumes 40%-60% fewer tokens than comparable models. This metric is more practically meaningful than raw benchmark scores — it means you can run more task cycles on the same compute budget.
Both models support a 1M token context window and are released in Base and Instruct variants.
Xiaomi’s AI Rhythm
Xiaomi hasn’t been the most aggressive player in large models, but this open source release sends several signals.
First, domestic teams’ pre-training capability is no longer just catching up. The Pro version’s SWE-bench Pro score matching Claude Opus 4.6 shows that in the code Agent vertical, open-source models can now compete head-to-head with closed-source frontier models.
Second, choosing MIT over stricter licenses like Apache 2.0 or commercial licensing lowers the psychological barrier for commercial adoption. For enterprises needing on-prem deployment with data staying in-house, this is a viable candidate for their evaluation shortlist.
Third, the simultaneous launch of the MiMo Orbit “100 Trillion Token Creator Incentive Plan” gives developers API quota — not cash, but compute. This mirrors the early cloud providers’ free tier strategy: get developers using it, build dependency, then figure out the business model.
Who Should Pay Attention, Who Can Wait
Try it now if:
- You’re building Agent systems and need an open-source baseline for comparison and iteration
- You have GPU resources (multi-card A100/H100) and need 1M long context
- You’re a startup that needs MIT license certainty for commercial freedom
Wait and observe if:
- Actual VRAM requirements and inference speed need verification — MoE architecture’s active parameters look low (42B/15B), but total parameters mean you need enough storage to load
- Chinese language capability — benchmarks are English-heavy; Chinese performance needs independent verification
- Whether Xiaomi will keep iterating — one open source release doesn’t build an ecosystem; update cadence matters
On the API side, Xiaomi’s MiMo API platform has both models live, with OpenCode among the first platforms to integrate.