MiniMax M3.1 Flash in practice: tests and comparisons with DeepSeek V4.1 Flash

A 300-task comparison finds Space Bunny (the anonymous M3.1 Flash) nearly identical to the final release with slightly better coding, roughly on par with DeepSeek but still behind on reasoning, with clear cost advantages for agent and enterprise workflows.

On the day MiniMax released M3.1 Flash weights, g2c published a brief deriving the model's design and architecture from first principles.

A developer reports these models are frontier-level for everyday tasks and swaps to M3.1 and others whenever Claude or GPT hits quota limits.

Official weights are ~252 GB, too big for two 256 GB DGX Sparks; the author sizes Mia's EXL3 3.06 bpw build at 167 GB (~84 GB per GPU under TP2) and estimates ~15.7 GB KV cache at 128K context, though no engine has verified loading yet.

The post notes M3.1 Flash ships 428B total parameters with 23B active per token and burned 100T tokens in three days, warning that open weights don't mean free local inference.

Two SAGE-EXL3 releases fit the 428B model on one or two DGX Sparks with 96.89% and 98.44% top-1 agreement, though vLLM currently caps context at 32K.

The 428B-parameter, 23B-active M3.1 Flash ships a roughly 252GB checkpoint, and within hours the community produced MLX 3-bit (~196GB), EXL3 (~171GB) and Q2_K GGUF (~160GB) variants, with day-zero vLLM NVFP4 support.

A demo game has three vision models, including MiniMax M3.1 Flash and DeepSeek V4.1 Flash, race to recognize doodles as they are drawn, all served on Nebius.

A breakdown of the M3.1 Flash release: 428B total / 23B active MoE, text-image-video input, 1M context, ~252GB 8-bit download with SGLang/vLLM/TokenSpeed recipes, free for non-commercial use but requiring attribution and revenue-based notice or approval commercially.

A developer says running DeepSeek 4.1-flash-1017 alongside MiniMax M3.1 Flash gets close to Opus 5.5 at a fraction of the price, calling both game changers.

One long prompt produced a working real-time three.js 3D forest shrine with volumetric fog, god rays, torii gates, falling petals, a ringing bell pavilion and cinematic drone camera paths, all followed in a single file; the author notes it is 3x faster than M3 at 1/3 the cost, $0.15 per million input tokens.

An article on M3.1 Flash's open weights covering 428B total / 23B active parameters, 1M context, vendor benchmarks running higher than outside numbers, a $20M clause in the license, and what to check before self-hosting.