Xiaomi, the Chinese electronics and electric vehicle maker, has released MiMo-V2.6-Pro, an open-weight AI model that now ranks as the world's best on independent benchmarks, surpassing DeepSeek and proprietary models from xAI and Google. The model scores 46 on Artificial Analysis' Intelligence Index, tying with xAI's newly released Grok 4.7 and beating DeepSeek V4.1 Pro (36) and DeepSeek V4.1 Flash (39). What makes this surprising is not just Xiaomi's unexpected entry into frontier AI, but the model's radical affordability and openness.

Why Should You Care About an Open-Weight Model?

Open-weight models are fundamentally different from proprietary ones like OpenAI's ChatGPT or Google's Gemini. Instead of renting access through an API, developers can download the model for free, customize it for their specific needs, run it on their own hardware, and deploy it in production without paying the company a cent. MiMo-V2.6-Pro is MIT-licensed, meaning it comes with minimal legal restrictions. For indie developers and enterprises alike, this represents genuine freedom.

The pricing advantage is striking. Xiaomi charges $0.435 per million input tokens and $0.87 per million output tokens through its API. By comparison, Artificial Analysis measured the model at $0.13 per task on its Intelligence Index benchmark. The company also released MiMo-V2.6-Flash, a smaller variant costing just $0.14 and $0.28 per million input and output tokens respectively, making it the second cheapest major frontier model available globally. Both models support a 1-million-token context window, meaning they can process roughly 750,000 words at once, with native support for text, images, audio, and video.

How Did Xiaomi Build a Model Better Than DeepSeek?

• Reinforcement Learning at Scale: Xiaomi invested millions in a technique called reinforcement learning (RL), where the model learns by attempting tasks, receiving rewards or penalties, and updating itself toward better behaviors. The company ran 30 large RL steps covering roughly 750,000 trajectories in under six days, costing approximately $2.62 million for the Pro model and $850,000 for Flash.
• Long-Horizon Agent Training: Rather than reinforcing short answers, Xiaomi reinforced entire lengthy agent workflows. Each training sequence averaged between 110,000 and 150,000 tokens, meaning the model learned from extended, complex reasoning chains involving planning, tool use, error correction, and strategy recovery.
• Diverse Training Environments: Xiaomi's "You Only RL Once" strategy mixed multiple domains and agent harnesses into a single training run, including coding, visual tasks, professional workflows, and cybersecurity. This diversity prevented the model from becoming tightly coupled to any single production setup.
• Efficient Compute Allocation: Of the Pro model's RL training cost, 43.5% went to updating the model itself, 43.8% to generating candidate solutions, and 12.7% to evaluating them. More than half the budget went toward generating and assessing the model's experience before that experience became weight updates.

What Makes This Different From DeepSeek's Approach?

DeepSeek's R1 models pioneered the use of chain-of-thought reasoning, where models show their working step-by-step before arriving at answers. This approach became influential across the industry, including in OpenAI's o-series models. However, Xiaomi's technical report reveals a different emphasis: rather than optimizing for visible reasoning transparency, the company optimized for clean, minimal solutions that solve problems efficiently. The report notes that simple binary rewards can tell a coding model whether its patch passes a test suite, but cannot reliably distinguish between two passing solutions where one is clean and minimal and the other introduces broad fallback logic. Xiaomi's training process appears designed to reward the former.

This distinction matters for real-world deployment. A model that learns to write minimal, maintainable code produces different outputs than one optimized purely for passing tests. For software engineers using these models in production, the difference between a clean solution and a hacky one can mean the difference between maintainable code and technical debt.

What Does Xiaomi's AI Ambition Actually Look Like?

The V2.6 release is not a standalone product launch but rather the latest step in Xiaomi's effort to build an entire open agent stack. The company began publicly expanding its MiMo family in 2025 and spent much of 2026 pushing deeper into agentic AI, which refers to AI systems that can plan and execute multi-step tasks autonomously. In April, MiMo-V2.5 and V2.5-Pro established the architectural and economic ideas now visible in V2.6: sparse mixture-of-experts models (which activate only relevant portions of their parameters), million-token context windows, permissive licensing, and unusually low prices for long-running agent workloads.

In June, Xiaomi released MiMo Code, an open-source terminal coding agent with persistent cross-session memory and task checkpoints. The company also introduced HarnessX, a research framework that treats the prompts, memory systems, tools, and control logic surrounding a model as components that can themselves be rewritten and optimized. Xiaomi reported an average 14.5% absolute performance gain across 15 model-benchmark combinations when the harness evolved dynamically, without replacing the underlying model. The V2.6 release pulls these lines of work back into the model-training process itself.

Xiaomi also released MiMo-V2.6-Pro-UltraSpeed, which the company says can generate output at up to 20 times the normal Pro speed, addressing latency-sensitive applications where inference speed matters more than raw reasoning capability.

What Does This Mean for the AI Industry?

Xiaomi's emergence as a frontier AI competitor signals a significant shift in the competitive landscape. The company is better known internationally for smartphones, electric vehicles, and consumer electronics than for artificial intelligence research. Yet it has deployed sufficient resources and technical expertise to build a model that outperforms both open-weight and proprietary alternatives on independent benchmarks. The combination of superior performance, open licensing, and aggressive pricing creates a compelling alternative for developers and enterprises that have traditionally relied on either expensive proprietary models or lower-performing open-source options.

The technical report's emphasis on reinforcement learning efficiency also suggests that scaling RL is becoming a viable path to model improvement, even for companies without the scale of OpenAI or Google. Xiaomi's reported costs of $2.62 million for Pro and $850,000 for Flash are substantial but not prohibitive for a company of Xiaomi's size, indicating that the frontier of AI capability may be broadening beyond the handful of US-based labs that have dominated recent headlines.