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

Unsloth published GGUF quants for M3.1 Flash: UD-IQ1_M at 129.4GB hits 89.9% top-1, UD-Q2_K_XL at 143.9GB 91.5%, UD-Q4_K_XL at 249.4GB 98.0%, with the 2-bit build needing 160GB+ RAM or VRAM.

MiniMax's official account highlights Unsloth's dynamic GGUF, which fits the 428B MoE into 143.9GB at 2-bit with 91.5% top-1 agreement across seven domains and runs via llama.cpp.

The author says MiniMax M3.1 Flash is one of the harder models to abliterate, noting there are no abliterated releases on Hugging Face while other models have many.

A user sizes up M3.1 Flash for Strix Halo: 428B total / 23B active and 235 GiB of weights mean even a 2-bit build is ~100 GiB before context, one class above what 128GB runs well.

Maya ships a dedicated engine for the 428B/23B-active M3.1 Flash, running locally on NVIDIA V100 and up with automatic VRAM planning, half-size long-context cache and lazy vision tower; GGUF quants for 128GB PCs are next.

Across 100 coding environments, M3.1 Flash lands at similar pricing to MiMo V2.6 Flash with lower performance, and the author says it is far from the Pareto front for open-weight or proprietary models.

An evaluator puts DeepSeek V4.1 Flash 1017 at 3.038, in a near-tie cluster with Meta Muse Wonder, Gemini 4 Argon, Mimo v2.6 Pro and MiniMax M3.1 Flash, strong at obstruction finding but weak on theorem scope.

Unsloth's MiniMax M3.1 Flash quants include a ~130GB 2-bit GGUF that is 45% smaller, retains 91.5% top-1 accuracy and runs on 150GB of memory.

Unsloth shipped UD3 quantizations of MiniMax-M3.1-Flash, with NVFP4 MoE layers by default and UD-IQ1_M retaining about 89.9% top-1; runs in Unsloth Desktop.

Unsloth's 2-bit GGUF shrinks the 428B MoE by 45% while retaining 91.5% top-1 accuracy, making local runs possible on 150GB of RAM.

On one video prompt, MiniMax M3.1 Flash cost $0.21 and produced bouncing debris, rising embers and trees burning one by one, while Kimi K3.1 Pro cost $6.10 for blocky cubes and solid lava — 29x the price for a worse result.

MiniMax released M3.1 Flash weights on Oct 23 with 428B total and 23B active parameters, sparse attention in the first three full-attention layers and W4A4 MoE experts, claiming 100T tokens processed in 72 hours and confirming it was the anonymous Space Bunny.