[REVISION HISTORY]

MiniMax AI model releases

2 recorded revisions of this situation.

What changed

2026-10-20 11:44 UTC → 2026-10-21 09:28 UTC · added ~~removed~~

AI developer MiniMax has released a series of specialized models focused on developer productivity and reasoning efficiency. In late September 2026, the company launched MiniMax-M3.1-Flash-Preview. This model was integrated into the MiniMax Code environment to assist with programming tasks such as bug localization and testing. It featured a 1-million-token context window and supported multimodal inputs including text, images, and video. By mid-October 2026, MiniMax released the M3.1 Flash reasoning model. This iteration aimed to reduce the costs of running AI agents by 70% compared to the previous M3 model, while increasing inference speeds up to 200 tokens per second. While the model outperformed competitors like Gemini 3.8 Flash and GLM-5.3-Flash in certain benchmarks, testing indicated a “high hallucination rate of 93%” because the model tended to attempt all questions rather than declining uncertain ones. On October 19, 2026, further details emerged regarding the M3.1 Flash model, which utilizes a mixture-of-experts (MoE) architecture. ~~The~~ Previously known under the anonymous pseudonym ‘Space Bunny’, the model features 428 billion total parameters. It has demonstrated strong performance in benchmarks for cybersecurity (CyberGym) and computer operation (OSWorld V2). MiniMax has positioned the model with aggressive pricing at $0.15 per million input tokens and $0.60 per million output tokens, and announced plans to release the model weights as open weights on October 23.

Versions

Select a version to see what changed compared to the one before it.