MiniMax's M3.1 Flash costs what DeepSeek charges for Flash only off-peak, so the contest between the two Flash models moves to benchmarks, speed and open-weight licensing.
MiniMax has turned the M3.1 Flash model it tested quietly in September into a full release, and has priced it to meet DeepSeek head-on. The Hong Kong-listed Shanghai company published the model on October 19 and said in its launch blog post that it charges RMB1 per million input tokens and RMB4 per million output tokens in China. Tech outlet Zhidongxi calculated that those rates are half of what DeepSeek charges for its own Flash model at peak hours, and equal to DeepSeek’s off-peak rates.
The release closes several gaps left open when EastFrontier covered the M3.1 Flash preview on September 28. At that point the model sat only in the model picker of MiniMax Code, the company’s coding assistant, with no published price, specifications or benchmarks. MiniMax now lists it on its developer platform with per-token pricing, has published benchmark scores and has said the model’s weights will be released for download later.
Every performance figure so far comes from MiniMax itself, and the company has not published a technical report. MiniMax also said a Pro version aimed at the highest capability level will follow, without giving a date.
MiniMax set its price two days after DeepSeek updated its Flash model
DeepSeek moved its Flash model to a new version, V4.1-Flash-1017, on October 17, according to China Fund News, and EastFrontier reported at the time that DeepSeek added Search and Fetch tools to that update. Zhidongxi listed DeepSeek’s peak rates as RMB2 per million uncached input tokens and RMB8 per million output tokens, with both halved off-peak.
MiniMax charges RMB1 per million input tokens, RMB4 per million output tokens and RMB0.1 per million tokens read from cache for M3.1 Flash. Its international pricing page lists US$0.15, US$0.60 and US$0.015 for the same three items, against US$0.30 and US$1.20 for input and output on its larger M3 model below 512,000 input tokens. China Fund News reported that input and output prices fell by more than half compared with M3, and cache reads by more than 70%.
MiniMax describes M3.1 Flash as a lightweight model built for coding and agent work, meaning software that carries out multistep tasks on a user’s behalf. The model can take in up to 1 million tokens of context. MiniMax’s English product guide describes a mixture-of-experts design, in which only part of the network runs for each request, with 428 billion parameters in total and about 23 billion active.
Speed claims and benchmark results vary by source
MiniMax says a more aggressive sparse attention design, which lets the model skip parts of a long input, makes online inference nearly three times faster. The company’s Chinese blog says everyday output exceeds 200 tokens per second, with a median above 170. The English product guide gives a lower figure of about 150 tokens per second and says actual speed depends on load and request length.
On benchmarks, MiniMax reported a score of 76.2 on BankerToolBench, an open test of agents handling investment-banking workflows, 73.7 on the DeepSWE v1.1 software engineering test, 61.7 on the OSWorld V2 computer-use test and 29.6 on Agents’ Last Exam. It described the last two as first among Chinese models.
The comparison with DeepSeek splits by task. Zhidongxi reported that DeepSeek’s new Flash model leads DeepSWE v1.1 with 77.0, ahead of M3.1 Flash, while MiniMax’s model scored 62.2 on Automation-Bench against 60.5 for DeepSeek. China Fund News reached a similar conclusion from public comparisons: DeepSeek is stronger on software engineering, while M3.1 Flash does better on cross-application business workflows and cybersecurity. MarkTechPost, a developer news site, noted that all of the scores are vendor-reported and that the test setup can move results.
There is also a technical difference. MarkTechPost reported that M3.1 Flash always reasons before answering, with five effort levels and no option to switch reasoning off, which sets a floor on cost and response time even for simple requests.
The weight release is the next deliverable
MiniMax said in its blog that the model’s weights will be published later. MarkTechPost reported a date of October 23 and said the license terms have not been disclosed. DeepSeek publishes its Flash weights under the MIT license, according to the same report, so the terms MiniMax chooses will shape how directly the two models compete for companies that run models on their own hardware.
The model is available now through the MiniMax API, the company’s M Plan subscription and the MiniMax Code desktop app, according to the blog. MiniMax also confirmed that Space Bunny, an anonymous model that Zhidongxi said at one point ranked first on the OpenRouter model marketplace, was an early build of M3.1 Flash and processed more than 200 trillion tokens before the launch.
The Hugging Face model card that accompanies the weights should settle the outstanding questions: the license, the hardware needed to run the model, and whether the speed and parameter figures match what MiniMax has stated so far.