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

vLLM announces Day-0 support for MiniMax-M3.1-Flash, with NVFP4 KV cache cutting memory 43.75% and roughly 2.7x more concurrent sequences on NVIDIA and AMD, plus notes on long-context cost and operations.

The author notes almost no one talks about MiniMax M3.1 Flash, guessing its plain MoE design leaves its accuracy-speed balance looking average next to Qwen3.8-Flash-Next and DeepSeek V4.1 Flash 1017.

The post cites 428B/23B active, 1M context and open weights, noting the official table shows 62.8 vs Opus 5's 50.3 on Automation-Bench but losses on Terminal Bench 4.0 and OSWorld 2.0.

After open-sourcing, M3.1 Flash costs about $0.22 per task versus several dollars for Claude and over three for GPT-6, but the author claims a 93% hallucination rate and fabrication instead of admitting ignorance.

M3.1 Flash weights are available in an October 23 repository, taking text, images and video and outputting text; the community license requires prominent attribution and prior written authorization above $20M yearly revenue.

A user shares a Hugging Face repo for DeRisked GGUF quantizations of MiniMax M3.1 Flash, expected to be uploaded the same day with no firm timing.

MiniMax released M3.1 Flash weights at 428B total parameters with only 23B active per token, downloadable on Hugging Face and ModelScope; the poster sees sparse activation keeping inference costs down.

A user says signing up on MiniMax's site and installing MiniMax Code yields 400 credits per day, and M3.1 Flash with 1M context and the harness works great.

Hands-on local run: oQ4e plus the DSpark drafter lifts one request from 22.5 to 42.0 tok/s (+87%) on an M3 Ultra, with 370 tok/s prefill at 4K and 298 tok/s at 128K.

A user bought the $11/month Go plan and installed MiniMax Code, praising its local/cloud toggle and phone remote control, and will decide after a month whether to switch to API access for M3.1 Flash.

A license comparison notes MiniMax M3.1 Flash has 428B parameters with 23B active and open weights, but its community license requires a "Built with MiniMax" credit and written approval for companies earning over $20M a year.

The poster ran M3.1 Flash (428B-A23B) on one 256GB M5 Ultra with TensorFold's native Metal engine, judged it not working well, guessed two DGX Sparks would do better, and shared an MLX repo.