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

MiniMax thanks the SGLang and Miles teams plus NVIDIA and AMD for supporting M3.1 Flash from inference to RL training on day one.

The author trained a DFlash2 speculative-decoding draft model for MiniMax M3.1 Flash, usable in SGLang from day 0, and notes Miles also has day-0 RL training support.

A user works around Antigravity's lack of custom model support, demoing with MiniMax M3.1 Flash, and notes remote PC control without using Gemini subscription quota.

The author says MiniMax M3.1 Flash improves notably over its predecessor but lacks competitiveness against DeepSeek V4.1 Flash 1017, with its $49 monthly usage quota covering most tasks as the bright spot.

SGLang adds day-0 serving for the open weights: over 1K tok/s at batch size 1, ~100K tok/s per GPU on the AgentX coding trace with 96 concurrent sessions, on Blackwell/Hopper and AMD MI350X.

NVIDIA announces day-0 fine-tuning support for the open-sourced MiniMax M3.1 Flash in NeMo-AutoModel, with FSDP2, expert/pipeline/context parallelism up to 128 GPUs and direct NVFP4 checkpoint loading.

HuggingPapers reports M3.1 Flash weights are on Hugging Face: a native multimodal MoE for coding, agents and computer use with 1M context and sparse attention.

A MiniMax staffer posted the Hugging Face and ModelScope links for downloading the open M3.1 Flash weights.

MiniMax released M3.1 Flash weights with 428B total and 23B activated parameters, plus agentic scores including 76.2 on BankerToolBench and 73.7% on DeepSWE v1.1.

A side-by-side video comparing MiniMax M3.1 Flash with Claude Opus 5.5, asking whether open source is catching up.

A short teaser saying MiniMax-M3.1-Flash open weights would be released the next day, with no further testing or details.

A video compares MiniMax M3.1 Flash with Claude Opus 5.5, with the author concluding Opus 5.5 remains hard to match.