Xiaomi Unveils MiMo-V2.6-Pro as World's Top Open-Weight AI Model Alongside Budget Flash Variant

Chinese consumer electronics and electric vehicle manufacturer Xiaomi has launched its MiMo-V2.6 model family, positioning MiMo-V2.6-Pro at the summit of third-party benchmarking firm Artificial Analysis Intelligence Index for open-weight systems. Scoring 46 on the index, the flagship model matches newly released proprietary offerings such as xAI's Grok 4.7 and surpasses several established closed and open models globally.

The release includes three core variants: the flagship MiMo-V2.6-Pro, the higher-throughput MiMo-V2.6-Pro-UltraSpeed, and the smaller MiMo-V2.6-Flash. Built as sparse mixture-of-experts architectures with native 1-million-token context windows supporting text, image, audio, and video inputs, the models are distributed under an MIT license permitting commercial utilization and self-hosting.

The technological core of the V2.6 generation centers on extensive reinforcement learning execution. According to Xiaomi's technical documentation, both the Pro and Flash checkpoints underwent 30 large-scale reinforcement learning steps covering roughly 750,000 trajectories over a period of under six days. The training expenses were reported at approximately $2.62 million for Pro and $850,000 for Flash.

The training regimen utilized fully asynchronous Group Relative Policy Optimization, processing 1,568 prompts per step with 16 candidate trajectories each, generating between 2.7 billion and 3.7 billion training tokens per iteration. Xiaomi implemented specialized mechanisms including Groupwise Reward Synthesis and Groupwise Advantage Redistribution to evaluate implementation quality, verify evidence gathering, and prevent reward-hacking behaviors such as code repository cloning during training cycles.

Beyond benchmark positions, the economic positioning of the MiMo series introduces competitive pressure across commercial API providers. Xiaomi maintains pricing structure from its V2.5 release, charging $0.435 per million uncached input tokens and $0.87 per million output tokens for the Pro flagship. The smaller MiMo-V2.6-Flash variant is priced significantly lower at $0.14 per million input tokens and $0.28 per million output tokens.

Artificial Analysis data highlights the price-performance positioning of the models, with Pro positioned on the intelligence-versus-cost Pareto frontier. Independent platforms and service providers, including OpenCode, have integrated the checkpoints into developer workflows, offering promotional access alongside self-hosted deployment options via Hugging Face and ModelScope.

The models demonstrate substantial capability gains over predecessor versions across long-horizon software engineering and agentic tasks. On the DeepSWE v1.1 benchmark, MiMo-V2.6-Pro achieved a score of 71.9, rising from 19.0 in V2.5-Pro. The Flash variant posted a score of 67.9, remaining competitive with larger architectures while operating at roughly one-third of the flagship's API cost.

MiMo-V2.6-Pro scores 53.1% on AutomationBench v1.0.6. Deploying open weights in domestic Indian data center regions guarantees data residency and air-gapped security for sensitive enterprise workloads.

Demonstrations accompanying the release showcased multi-agent coordination across 3D scene generation, Blender asset creation, robotic arm control via multi-view visual feedback, and scientific research assistance. Xiaomi's materials researchers utilized Pro to evaluate metal-organic frameworks for capturing chemical pollutants, while separate academic validations verified automated Lean 4 code formalizations exceeding 6,000 lines.

Xiaomi is releasing more than raw model weights, providing the technical report, over 7,000 reinforcement learning task environments, composable mini-harnesses, and a distilled MiMo-V2.6-Distill-Qwen-9B model. Researchers can inspect training curves, grader configurations, and reward mechanisms to evaluate reproducibility. Enterprise adoption hurdles remain centered on self-hosting infrastructure requirements for the trillion-parameter Pro checkpoint and regulatory considerations regarding API server jurisdictions.

The 9B distilled model advances coding benchmarks over base Qwen3.5-9B, increasing SWE-bench Verified from 60.0% to 61.1% and SWE Pro from 32.0% to 44.6%. MiMo-V2.6-Pro retains the 1.02T total parameter and 42B active parameter MoE backbone from V2.5 while incorporating a 1M-token context window, hybrid SWA/GA attention, and a five-layer speculative decoding drafter.