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

Weights are public on Hugging Face and ModelScope: 428B total / 23B active MoE with 1M context, DeepSWE 73.7%, CyberGym 82.2, OSWorld V2 61.7, API at $0.15/$0.60, roughly a third cheaper and 3x faster than M3.

A user has Codex offload deterministic skills to mcode running M3.1 Flash with GPT as reviewer, already producing Remotion videos this way and reporting roughly half the quota used.

A post says DeepSeek and MiniMax obtained open-source permissions for V4.1F-1017 and M3.1-Flash respectively under new Chinese regulations.

MiniMax released M3.1 Flash weights on Hugging Face and ModelScope: a 428B MoE with 23B active, sparse attention and speculative decoding for ~3x speed, DeepSWE 73.7%, commercial license needed above $20M revenue.

An NVIDIA-side guide shows RL fine-tuning MiniMax M3.1 Flash with NeMo RL: GRPO on the AutoModel backend, context plus expert parallelism (CP2 + EP128) and vLLM generation.

Analysis of the M3.1 Flash open-weight release: 428B total / 23B active, official claim of over 100 trillion tokens processed by real users in 72 hours, and same-day MetaX C-series GPU adaptation, though benchmark comparisons, token-count methodology and post-adaptation throughput lack third-party data.

MiniMax M3.1 Flash open weights are out at 428B total parameters with 23B active per token, and the poster calls its agent benchmark numbers impressive on paper.

Weights are downloadable on Hugging Face and ModelScope with 428B total and 23B active parameters per token, released on the promised 23rd.

MiniMax-M3.1-Flash is on Hugging Face with 428B total and 23B active parameters, matching M3, and the poster finds its capability solid for that size.

A vLLM blog post covers day-0 MiniMax M3.1 Flash support with optimized NVIDIA and AMD recipes and AgentX Pareto curves, described as the most comprehensive Rubin day-0 support yet.

A user notes that MiniMax M3.1 Flash, previously the anonymous Space Bunny, can now be run locally, without sharing hardware or speed details.

ModelScope lists M3.1 Flash's 428B MoE with 23B active, 1M context and native image/video input, plus official scores like 62.8 Automation-Bench, 73.7% DeepSWE, 61.7 OSWorld 2.0 and 82.2 CyberGym, with SGLang/vLLM self-hosting guidance.