Xiaomi has released MiMo-V2.6-Pro, an open-weight artificial intelligence model that ranks as the world's best-performing freely available AI system, tying with xAI's Grok 4.7 on independent benchmarks while costing significantly less to use. The model scores 46 on Artificial Analysis' Intelligence Index, surpassing Google's Gemini 3.8 Flash (41) and DeepSeek's latest versions. Unlike proprietary competitors, developers can download, customize, and run MiMo-V2.6-Pro for free or access it through Xiaomi's API at rates among the cheapest in the industry.

What Makes Xiaomi's New Model Different From ChatGPT and Other Competitors?

The distinction between MiMo-V2.6-Pro and closed-source models like OpenAI's GPT-4 comes down to licensing and cost. MiMo-V2.6-Pro carries an MIT license, meaning developers worldwide can download it from Hugging Face, fine-tune it for specific tasks, and deploy it on their own hardware without paying Xiaomi anything. For those who prefer cloud access, Xiaomi charges $0.435 per million input tokens and $0.87 per million output tokens, making it substantially cheaper than most enterprise-grade alternatives at its performance level.

The model processes roughly 100,000 words at once, supports text, images, audio, and video inputs, and generates responses at approximately 134 tokens per second. Xiaomi also released MiMo-V2.6-Flash, a smaller variant priced at $0.14 and $0.28 per million input and output tokens respectively, designed for high-volume production workloads.

How Did Xiaomi Build a Model That Competes With Industry Leaders?

The technical foundation behind MiMo-V2.6-Pro centers on reinforcement learning, a training method where models improve by attempting tasks, receiving feedback on performance, and adjusting their behavior toward better outcomes. Xiaomi invested millions scaling this approach, spending approximately $2.62 million to train MiMo-V2.6-Pro and $850,000 for the Flash variant. Each training step involved 1,568 prompts generating 16 candidate solutions each, producing roughly 25,000 rollouts and between 2.7 billion and 3.7 billion training tokens per step.

What distinguishes Xiaomi's approach is where the budget went. The company allocated 43.5% of training costs to updating the model itself, 43.8% to generating candidate solutions, and 12.7% to evaluating which solutions deserved reinforcement. This distribution reflects Xiaomi's focus on reinforcing lengthy, multi-step workflows rather than short answers. The resulting sequences averaged between 110,000 and 150,000 tokens each, simulating real-world agent tasks that span hundreds of tool calls.

Steps to Understanding Xiaomi's Technical Innovation

• Reinforcement Learning at Scale: Xiaomi completed 30 large reinforcement learning steps covering approximately 750,000 trajectories in under six days, teaching the model to improve through repeated practice on complex tasks rather than static examples alone.
• Multi-Domain Training: Instead of running separate reinforcement learning programs for coding, visual work, computer use, and cybersecurity, Xiaomi mixed all domains into a single training run, exposing the model to diverse agent harnesses and system prompts.
• Asynchronous Processing: The model uses fully asynchronous Group Relative Policy Optimization, allowing long-running agent jobs to finish at different times without stalling the entire training process, keeping computing hardware continuously busy.
• Nuanced Reward Systems: Rather than simple pass-fail grading, Xiaomi's reward system distinguishes between solutions that merely pass tests and those that solve problems cleanly and efficiently, encouraging elegant code over brute-force approaches.

Why Should Enterprises and Developers Care About This Release?

MiMo-V2.6-Pro represents a shift in how frontier artificial intelligence becomes accessible. Enterprises with data privacy concerns or restrictions on using Chinese-based servers can download the model and run it entirely on their own infrastructure, eliminating reliance on external APIs. Indie developers gain access to a top-tier model without subscription costs, enabling experimentation and customization that proprietary systems restrict.

The pricing structure also matters for cost-sensitive workflows. At $0.13 per Intelligence Index task according to Artificial Analysis' measurement, MiMo-V2.6-Pro delivers performance comparable to much more expensive alternatives. The Flash variant, priced even lower, targets organizations running high-volume production workloads where cost per inference becomes critical.

Xiaomi's broader strategy extends beyond a single model release. The company has spent 2026 building what it calls an "open agent stack," including foundation models, coding agents, training frameworks, and reinforcement learning environments. MiMo Code, released in June, functions as an open-source terminal coding agent with persistent memory across sessions. HarnessX, Xiaomi's research framework, treats prompts, memory systems, tools, and control logic as optimizable components, reporting an average 14.5% absolute performance gain across 15 model-benchmark combinations when the harness evolved dynamically.

The timing of MiMo-V2.6-Pro's release alongside xAI's Grok 4.7 underscores intensifying competition in the frontier AI space. Both models scored 46 on Artificial Analysis' latest methodology, yet they represent different philosophies: Grok 4.7 remains proprietary and closed, while MiMo-V2.6-Pro is open-weight and permissively licensed. This divergence reflects a broader industry trend where open-source and proprietary approaches coexist, each serving different customer needs and risk profiles.