Xiaomi Releases Open-Weight Multimodal Model MiMo-V2.6-Pro… Higher Intelligence and Competitive Pricing Compared to Similar-Sized Models

Xiaomi has released the open-weight multimodal model MiMo-V2.6-Pro. This model adopts a Mixture of Experts (MoE) architecture with 1.0T total parameters and 42B active parameters, supports a 1M tokens context window, and accepts text, image, speech, and video inputs. The license is MIT, allowing commercial use and self-hosting.

Performance and Intelligence Evaluation

It scored 46 points on the Artificial Analysis Intelligence Index. This is a high score compared to the median (18 points) of similar-sized open-weight models, reflecting a comprehensive evaluation of reasoning, knowledge, mathematics, and coding. The original text evaluated this model as one of the 'leading models in intelligence'.

Cost Efficiency

Price competitiveness is noteworthy. Input costs are set at $0.43/1M tokens, and output costs at $0.87/1M tokens. According to the original text, the output price is evaluated as 'very competitive' compared to the median of similar-sized models ($1.68). The input price is mentioned as either better than the median ($0.45) or slightly more expensive ($0.30) depending on the context within the original text, but is overall evaluated as a reasonable price range. The total evaluation cost was $206.66.

Speed and Limitations

The output speed is 57.9 tokens/sec, which is slower than the median (67.7 t/s) of similar-sized open-weight models. Latency (TTFT) was also measured at 2.61s, slightly higher than the median (2.34s), indicating it is inferior to the average in terms of speed. It is currently available via Xiaomi's API.