MiniMax M3.1 Flash
A lightweight, high-speed text model released by MiniMax, targeting everyday development scenarios. It is designed for low latency and high response speed, making it suitable for high-frequency tasks such as rapid code fixes and similar lightweight workloads.
Core Facts
MiniMax M3.1 Flash Preview emphasizes speed and can be used for quickly fixing bugs in code development.
MiniMax M3.1 Flash is available on AIHubMix for token-billed calls under the model ID `minimax-m3.1-flash`, with input priced at $0.165 per million tokens and output at $0.66 per million tokens.
Calculated per thousand repair tasks, Claude Haiku 5.5 costs about $0.70, DeepSeek V4.1 Flash 1017 about $1.35, and MiniMax M3.1 Flash about $4.45 by default and about $0.64 at the low setting.
Timeline (last 90 days)
DeepSeek V4.1 Flash (the October 17 update referred to as 1017) and MiniMax M3.1 Flash official version were compared and tested on real archives exceeding 100,000 tokens, the official APIs, and the same AI Max hardware.
In the low-tier recall sample, DeepSeek output 4836 tokens and MiniMax output 1487 tokens, both passed the three functional checks, with wait times of about 23 seconds and 20 seconds.
In the speed test, dropping simple filtering from the Max tier to the Low tier did not save DeepSeek time; MiniMax's median fell from 5.2 seconds to 2.1 seconds, and both were correct in two rounds at the Low tier.
In the amount calculation and refund triage question on archives exceeding 100,000 tokens, DeepSeek got it fully right in one round with cited evidence, and MiniMax official version got the core amounts right in both rounds but did not pass the strict score, with problems in literal formatting such as evidence prefixes and version labels.
For strict delivery across the whole task, all three models were 0 out of 2 (failing both rounds), and final file correctness was 1 out of 2 (one correct round out of two) for all.
In the release step (with an injected disk-full fault), DeepSeek and MiniMax official version each completed recovery in one round and passed 35 functional checks, but both stopped at the missing stage wrap-up; MiniMax Preview completed even the wrap-up in one round.
In the decision memo question, none of the three models fully met the standard in either of the two rounds, with a threshold of 16 points; DeepSeek produced no finished draft in both rounds, MiniMax official version wrote half a sentence in one round and scored 14 points on editing in the other, and the old Preview scored 13 and 12 points.
In the scheduling function question (handling dependent resources and exceptions), DeepSeek official and the old Preview each passed one round, while MiniMax official version exhausted more than eight thousand output tokens in both rounds without passing.
In the round where MiniMax official version failed the inventory question, its actual actions were correct, but it added a code fence outside the final JSON, causing strict interface parsing to fail.
In the inventory question (select the right product and reserve only once), DeepSeek completed it in both rounds; MiniMax official version and Preview each passed one round.
40 more timeline events
All Facts (20)
MiniMax M3.1 Flash Preview emphasizes speed and can be used for quickly fixing bugs in code development.
65%Partially VerifiedMiniMax M3.1 Flash is available on AIHubMix for token-billed calls under the model ID `minimax-m3.1-flash`, with input priced at $0.165 per million tokens and output at $0.66 per million tokens.
65%Partially VerifiedCalculated per thousand repair tasks, Claude Haiku 5.5 costs about $0.70, DeepSeek V4.1 Flash 1017 about $1.35, and MiniMax M3.1 Flash about $4.45 by default and about $0.64 at the low setting.
65%Partially VerifiedMaximum output tokens: MiniMax M3.1 Flash is 512K, DeepSeek V4.1 Flash 1017 is 384K, and Claude Haiku 5.5 is 128K.
65%UnverifiedMiniMax M3.1 Flash uses a MoE architecture with approximately 428B total parameters and approximately 23B activated per token.
60%UnverifiedM3.1 Flash can be called via the AI SDK, Chat Completions, Responses, and the Anthropic Messages API, with the model name minimax/minimax-m3.1-flash.
60%UnverifiedMiniMax M3.1 Flash offers a 1M-token ultra-long context window.
60%UnverifiedThe official name of MiniMax M3.1 Flash is `MiniMax-M3.1-Flash`, positioned as a frontier multimodal coding model with a 1 million context window and adjustable thinking depth.
60%UnverifiedIn the dual-bug Python function repair benchmark (215 hidden tests), omitting effort (max) consumes an average of 6,364 reasoning tokens, takes 56 seconds, and costs $0.0045 per run; effort=low consumes an average of 591 tokens, takes 9.5 seconds, and costs $0.0006 per run.
60%UnverifiedMiniMax M3.1 Flash defaults to the max setting when the effort field is not passed in the request, and the model always performs thinking.
60%UnverifiedWhen the reasoning_effort field is not passed, MiniMax M3.1 Flash runs at the max thinking level, taking an average of 56.3 seconds for a small repair task, which may trigger client timeouts tuned for fast models.
60%UnverifiedOn September 27, MiniMax named Space Bunny Alpha M3.1-Flash-Preview and made it the default model for MiniMax Code.
60%UnverifiedWait times for the basic rule question: DeepSeek official median about one second, DeepSeek official version about five seconds, MiniMax Preview about 1.5 seconds.
50%UnverifiedIn the inventory question (select the right product and reserve only once), DeepSeek completed it in both rounds; MiniMax official version and Preview each passed one round.
50%UnverifiedDeepSeek V4.1 Flash (the October 17 update referred to as 1017) and MiniMax M3.1 Flash official version were compared and tested on real archives exceeding 100,000 tokens, the official APIs, and the same AI Max hardware.
50%UnverifiedIn the scheduling function question (handling dependent resources and exceptions), DeepSeek official and the old Preview each passed one round, while MiniMax official version exhausted more than eight thousand output tokens in both rounds without passing.
50%UnverifiedIn the basic rule question (filter illegal data and must not modify the input), all three models passed the nine functional checks in two rounds, with no gap in quality.
50%UnverifiedIn the round where MiniMax official version failed the inventory question, its actual actions were correct, but it added a code fence outside the final JSON, causing strict interface parsing to fail.
50%UnverifiedIn the decision memo question, none of the three models fully met the standard in either of the two rounds, with a threshold of 16 points; DeepSeek produced no finished draft in both rounds, MiniMax official version wrote half a sentence in one round and scored 14 points on editing in the other, and the old Preview scored 13 and 12 points.
50%UnverifiedIn the release step (with an injected disk-full fault), DeepSeek and MiniMax official version each completed recovery in one round and passed 35 functional checks, but both stopped at the missing stage wrap-up; MiniMax Preview completed even the wrap-up in one round.
50%