Updates·October 19, 2026, 14:37

Cheap Chinese AI model beats rivals on select tests but often makes up answers

AI-generated from the sources listed, with automated source checks.

MiniMax has released the language model M3.1 Flash, which is cheaper and faster than its predecessor. But the model has a hallucination rate of 93 percent, and the open weights are not planned until October 23.

Chinese AI company MiniMax released the language model M3.1 Flash on October 19, 2026. According to The Decoder, it costs $0.15 per million input tokens and $0.60 per million output tokens. A token is a small chunk of text that the model reads or writes. By comparison, the earlier M3 costs $0.30 for input and $1.20 for output, twice as much.

First anonymous, then revealed

Before the launch, the model was hidden on the OpenRouter service under the name "Space Bunny." According to VentureBeat, it processed more than 200 billion tokens before MiniMax revealed that it was theirs. A preview version appeared as early as September 27, but it could only be used through a subscription plan and the MiniMax Code tool, not with pay-per-token billing.

Faster and cheaper

The Decoder writes that the model is three times faster than M3. VentureBeat reports a speed of over 200 tokens per second. The model takes in text and images and responds with text. It can hold 1 million tokens in memory at once.

The testing firm Artificial Analysis gave the model 42 points on its Intelligence Index, compared with 29 for M3. One task cost about $0.22, less than half the cost for M3.

Against rival Deepseek V4.1 Flash, MiniMax scores higher on factual accuracy (56 versus 43 percent) and on the difficult Humanity's Last Exam (48 versus 43 percent). Deepseek leads on the Intelligence Index (45 versus 42), however, and MiniMax is slightly cheaper per task ($0.22 versus $0.26).

The weakness: It guesses rather than hesitates

The model answered correctly on 56 percent of the factual questions in the AA-Omniscience test. But its hallucination rate is 93 percent. That means it almost never says "I don't know" when it is wrong, and instead gives an incorrect answer. For everyday users, that is a reason to double-check important answers.

What comes next?

According to VentureBeat, MiniMax plans to release the model's open weights on October 23 on Hugging Face. With open weights, others can download and run the model itself. The license terms have not yet been published. According to VentureBeat, a more powerful Pro model is also on the way.

The API is live at platform.minimax.io, and the model can also be used through the MiniMax Code app. The figures from MiniMax's own tests, such as 73.7 percent on DeepSWE v1.1, are the company's own measurements.