Xiaomi introduces MiMo-V2.6 Pro with striking performance

Xiaomi has launched its latest AI model this week: the MiMo-V2.6 Pro, which according to the Intelligence Index of Artificial Analysis scores the highest among open models. With 46 points, the Pro model outperforms competitors such as Kimi K3 and Qwen. The model employs a mixture-of-experts architecture with 1.02 trillion parameters, utilizing 42 billion active parameters for each request. Additionally, a smaller variant, the MiMo-V2.6-Flash, is also being introduced.

The MiMo-V2.6 Pro also supports higher performance on the DeepSWE coding test, with a score rising from 58.4 to 72.6, while the Flash variant increases from 48.8 to 65.7. Furthermore, Xiaomi introduces the Pro-UltraSpeed, which is capable of delivering executions up to 20 times faster. In terms of costs, the price for processing tokens is on the low side, at €0.38 per million input tokens and €0.76 per million output tokens. A test task costs about €0.11, enhancing the appeal of the MiMo-V2.6 Pro.

The impressive performance of the model is attributed to an improved approach to reinforcement learning, with Xiaomi focusing on more data, task variation, and increased computing power. This technological advancement led to substantial investments, amounting to approximately $2.62 million for the Pro model and $0.85 million for the Flash variant. Additionally, Xiaomi has made its RL toolkit public, including a technical report, training framework, and an offering of 7,000 ready-to-use training tasks covering various applications, from software development to music composition.

However, the launch of the MiMo-V2.6 Pro comes at a sensitive time, as Anthropic accuses Xiaomi of abusing Claude. A recent report mentions seven Chinese labs, including Xiaomi, allegedly involved in illegal activities to mimic Claude's capabilities. This report highlights the complexity of the competitive landscape within the AI sector and casts a shadow over the remarkable performance of Xiaomi's new models.

Read the full article from AI Insider.