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

Unsloth's 2-bit dynamic GGUF shrinks the 428B M3.1 Flash to 143.9GB, runnable on a 150GB RAM machine with 91.5% top-1 agreement versus the original, plus official DeepSWE/OSWorld/CyberGym scores.

Using Unsloth's 2-bit GGUF (144GB) across 96GB VRAM plus system RAM, a custom forked engine hit 23–31 tok/s decode and 1120 tok/s prefill at 16K versus llama.cpp's 6.7 and 219 tok/s, with identical NLL and RAM bandwidth as the bottleneck.

A user says DeepSeek won the comparison, but both MiniMax M3.1 Flash and DeepSeek V4.1 Flash 1017 are cool and neither run was perfect; MiniMax feels more fun and they plan to test its code cloud and remote control.

Two DGX Sparks run the 428B M3.1 Flash at 2.15 bpw, hitting 33 tok/s on rewriting and 29 on code with 13.6s for an 8k prompt, but IFEval drops to 60.5% versus 89.0% for DeepSeek V4.1 Flash at 2.9 bpw, so a 3.06 bpw pack is being tested.

The author recommends M3.1 Flash for boring background agent jobs but not near production code without checking, citing $0.15 input, $0.60 output and $0.015 cached input per 1M tokens.

Unsloth's 2-bit dynamic GGUF shrinks the 428B MoE MiniMax M3.1 Flash to about 144GB with 91.5% top-1 token agreement in its own eval, runnable on 150GB of RAM.

Hands-on: one prompt yielded a playable single-file rain-soaked city runner in 46 minutes; self-playtesting found the runner falling through stacked rooftops after a few hundred meters, fixed with a 30,000-step check passing 96/96 with zero console errors.

Local inference project Project Maya now has the MiniMax-M3.1-Flash DSpark drafter read the whole conversation, making answers up to 1.5x faster with identical output, alongside a new web UI and AMD/MI50 support.

An independent security benchmark found M3.1 Flash hit 24/32 CVEs at pass@3 (75%) vs DeepSeek V4.1 Flash's 28/32 (87.5%), with 59.4% vs 70.8% single-run recall, ~40% more false positives, and $4.20 vs $2.40 per CVE; the author does not recommend it for security work.

A hands-on test has M3.1 Flash build a multi-file 3D candy factory game in OpenCode — six machines, belts, economy, live charts, minimap and save/load — fixing its own bugs along the way, and it was fast.

Hands-on: given one prompt for a playable pixel adventure game, M3.1 Flash in MiniMax Code built it, playtested it in the built-in browser, iterated on gameplay and UI, and caught its own bugs; 400 free credits daily.

The author ran M3.1 Flash via API in DeepSeek Harness and mcode over long sessions, adding features one by one to build a code-only open-world FPS with no external assets, calling it a big step up from M3 and planning to spend the rest of the month's quota improving it.