MiniMax M3
A large language model released by MiniMax, featuring extended long-context processing capabilities at the million-token scale and native multimodal understanding. The model supports unified reasoning across text, images, and video.
Timeline (last 90 days)
Testing the same scene and the same sentence 10 times each, DeepSeek's average response time was 3.68 seconds, M3's was 10.70 seconds, and pre-optimization M3.1's was as high as 20.35 seconds.
Between the two backup models, M3.1 was preferred over M3.
Three models—DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3—were released on the same day.
MiniMax M3 allows reasoning to be disabled, whereas M3.1 Flash does not.
MiniMax M3 is built on MiniMax Sparse Attention, making the operating cost of a 1 million token context far lower than the previous generation.
The community matched Space Bunny Alpha to MiniMax through tokenizer fingerprinting: X user cheaty found that its text tokenizer is exactly identical to MiniMax M3's, and other testers cross-validated this conclusion with 36/36 and 50/50 probe strings.
On AIHubMix, MiniMax M3.1 Flash's input price is about 44% lower than M2.7's ($0.165 vs. $0.2958) and about 43% lower than M3's ($0.165 vs. $0.288), with the same ratios on the output side.
The model catalog refresh added two providers, Boundless and Standard Compute; SambaNova is now fetched from the catalog, with MiniMax M3 as the default model.
The Cline recommendation list adds Gemini 4 Argon and removes Grok 4.7; the free list replaces Step 5 Preview with Nemotron 3.5 Super VL; OpenCode Go no longer lists MiniMax M2.7.
SambaNova is now fetched from the catalog, with a default model of MiniMax M3.
8 more timeline events
All Facts (16)
Testing the same scene and the same sentence 10 times each, DeepSeek's average response time was 3.68 seconds, M3's was 10.70 seconds, and pre-optimization M3.1's was as high as 20.35 seconds.
50%UnverifiedBetween the two backup models, M3.1 was preferred over M3.
50%UnverifiedMiniMax M3 allows reasoning to be disabled, whereas M3.1 Flash does not.
50%UnverifiedMiniMax M3 is built on MiniMax Sparse Attention, making the operating cost of a 1 million token context far lower than the previous generation.
50%UnverifiedOn AIHubMix, MiniMax M3.1 Flash's input price is about 44% lower than M2.7's ($0.165 vs. $0.2958) and about 43% lower than M3's ($0.165 vs. $0.288), with the same ratios on the output side.
50%UnverifiedThe community matched Space Bunny Alpha to MiniMax through tokenizer fingerprinting: X user cheaty found that its text tokenizer is exactly identical to MiniMax M3's, and other testers cross-validated this conclusion with 36/36 and 50/50 probe strings.
50%UnverifiedSambaNova is now fetched from the catalog, with a default model of MiniMax M3.
50%UnverifiedThe Cline recommendation list adds Gemini 4 Argon and removes Grok 4.7; the free list replaces Step 5 Preview with Nemotron 3.5 Super VL; OpenCode Go no longer lists MiniMax M2.7.
50%UnverifiedBottlenecks in inference serving can occur at multiple stages, including computation, memory bandwidth, scheduling, and communication.
50%UnverifiedMiniMax M3 employs complex architectures such as Mixture of Experts (MoE), imposing higher demands on sparse computation and large-scale KV Cache management.
50%UnverifiedOptimizing MiniMax M3 inference serving on the MI355X ultimately achieved a 3–4x improvement in serving throughput.
50%UnverifiedGen-1 Slides was obtained through post-training on MiniMax AI's M3 model as the base.
50%Unverified测试采用控制变量法，将同一份提示词交给内置MiniMax M3模型的MiniMax Code CLI和手动接入MiniMax模型的Claude Code，以只比较工具链差异
50%UnverifiedThe model catalog refresh added two providers, Boundless and Standard Compute; SambaNova is now fetched from the catalog, with MiniMax M3 as the default model.
70%UnverifiedThree models—DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3—were released on the same day.
50%UnverifiedAMD and EmbeddedLLM jointly published a blog documenting the process of optimizing MiniMax M3 inference serving on Instinct MI355X hardware.
50%