Minimax launches M3.1 Flash with high hallucination rate

Minimax released M3.1 Flash, a new reasoning model that scores 13 points higher than its predecessor on the Artificial Analysis index and costs less than half per task. It ranks slightly ahead of Gemini 3.8 Flash and behind DeepSeq V4.1 Flash, though it outperforms DeepSeq in complex reasoning and factual fidelity due to lower output token usage. The model supports up to one million tokens of context and images and is available via API, with model weights to follow. A significant weakness is a 93% hallucination rate, as the model attempts to answer almost every question rather than declining when uncertain.

Data from the Versils AI Gateway Index shows that over 70% of data traffic now runs through models with publicly accessible weights, though their revenue share is lower. In September, nearly half of teams switched their most-used model, and a third switched providers, indicating AI models are becoming interchangeable components. The high hallucination rate of M3.1 Flash highlights the trade-off between cost efficiency and reliability in fast reasoning models. Users must carefully evaluate the model's suitability for tasks where factual accuracy is critical. The release underscores the rapid iteration and commoditization of AI models in the market.