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

MiniMax released M3.1 Flash weights on Hugging Face: a 428B MoE activating 23B per token with 1M context and image/video input, about 252GB at 4-bit (MLX 248GB at 4-bit, 187GB at 3-bit), fitting a 512GB Mac Studio or two linked DGX Sparks. Self-reported DeepSWE 73.7 vs Opus 5's 74.0, but behind on Terminal-Bench (38.1 vs 51.8) and OSWorld 2.0 (61.7 vs 70.6); commercial use needs attribution and pr

Five head-to-head tests on identical prompts, including MiniMax M3.1 Flash at $0.21 versus GPT-6 Astra at $7.88 for the same ice rink, plus Gemini 4, Opus 5.5 and Seedance comparisons.

A developer built Repro: it records console, network and clicks, has M3.1 Flash on Nebius write the bug report, reproduces in the signed-in browser, proposes a fix and verifies it in Docker; a real store bug reproduced in 2 steps with a 1-file fix.

The author learned the mystery model that built their fluid simulation, Space Bunny, is MiniMax M3.1 Flash; handing back a 1,564-line single file, it split it into 13 files in 18.5 minutes for $0.17 at API prices, after topping OpenCode for 12 anonymous days with 106T tokens served.

Msty's account posts a video claiming MiniMax M3.1 Flash can now run on your own local cluster, without VRAM, quantization or throughput details.

MiniMax published M3.1 Flash weights on Hugging Face, and the post notes the model's design strength and that it can now be run locally.

A Japanese user briefly says MiniMax M3.1 Flash suits their taste better than DeepSeek V4.1 Flash 1017, without citing tasks or numbers.

The poster notes MiniMax open-sourced M3.1 Flash weights with local-run and million-token context claims, then asks whether to self-host or stick to the API, with no hands-on test.

Given a prompt about a little astronaut bunny collecting fallen stars on a floating moon island, MiniMax M3.1 Flash built a playable 3D game, Moonbloom, through OpenCode with $0 spent on model usage.

A user had DeepSeek V4.1 Flash write code, then had MiniMax M3.1 Flash and GPT-6.1 Sol review it, and the review turned up a large number of vulnerabilities.

A user says the 1TB drive on their gx10 was enough before, but DeepSeek V4.1 Flash's ~470GB weights plus engram and Docker images filled it, leaving no space to copy M3.1 Flash weights for local testing, so an SSD upgrade may be needed.

Using three subagents (two reading code, one verifying), the author screenshotted old features and wrote passing Playwright tests as an answer key, split the file piece by piece committing only on green tests, then ran jscpd to dedupe, knip to delete dead code, and moved hard-coded keys to a gitignored .env with rotation; about 17 cents at API prices in 18.5 minutes.