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

A hands-on comparison on a real AI-video workflow: M3.1 Flash delivered the storyboard tool's first draft in 6m36s vs 8m39s and revisions in 2m13s vs 3m17s, kept prior edits intact, and proposed a more striking fight choreography.

Weights are on Hugging Face with 428B total and 23B active parameters, 1M context, native multimodality, 3x speed and 70% lower agent cost under a community license.

One prompt produced FRACTURE, a 6,000-particle interactive simulation with real-time physics and controls; the author notes 3x M3 speed, $0.15/M input pricing and open weights.

The author plans to fine-tune baked-in guardrails away for local vulnerability testing, saying DeepSeek V4.1 Flash 1017 tends to cause thinking loops while M3.1 Flash seems safer, and has started tuning it.

For a 30-second 5-scene 3D supercar video, GPT-6.1 Sol finished in 4m51s for $0.14 on one try, while M3.1 Flash took 14m15s and $0.058 over two tries; GPT looked more real, MiniMax won on value.

One prompt produced a one-minute Three.js Dhaka city vlog with four scenes, camera moves, captions and a weather change, with no image or model assets, while the model worked autonomously for 3h29m.

A user cites someone else's evaluation saying Space Bunny, the anonymous M3.1 Flash, is roughly on par with DeepSeek 4.1 Flash 1017 with different strengths.

Local deployment report: the 428B/23B-active M3.1 Flash runs on 4x DGX Spark with a day-0 vLLM TP4 recipe at ~40 tok/s prose and ~62 tok/s structured at c1, 262K context self-reported, with 2-Spark EXL3 and pruned NVFP4 options coming.

Local-deployment notes: 428B MoE with 23B active, 1M context, multimodal, ~250GB QAT checkpoint; Artificial Analysis score 42 ties GLM-5.3-Flash. Two DGX Sparks are too tight with context, four at tp4 fit comfortably, no Spark benchmarks yet.

The post notes MiniMax M3.1 Flash includes vLLM and SGLang recipes with paged attention and continuous batching out of the box, but its community license bars commercial use without permission, suiting research or teams willing to negotiate.

A user shares the Hugging Face link confirming MiniMax M3.1 Flash weights are out, joking about Halloween costume ideas for them; no benchmarks or specs in the post.

A free, commercially usable draft model for M3.1 Flash that keeps output identical while claiming up to 3.5x speedup; measured 3.53x on a single request with two GB300s, dropping to 1.6–2.6x at 8–32 concurrent requests.