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

Unsloth's 2-bit dynamic GGUF shrinks the 428B M3.1 Flash to 143.9GB, runnable on a 150GB RAM machine with 91.5% top-1 agreement versus the original, plus official DeepSWE/OSWorld/CyberGym scores.

Two DGX Sparks run the 428B M3.1 Flash at 2.15 bpw, hitting 33 tok/s on rewriting and 29 on code with 13.6s for an 8k prompt, but IFEval drops to 60.5% versus 89.0% for DeepSeek V4.1 Flash at 2.9 bpw, so a 3.06 bpw pack is being tested.

An independent security benchmark found M3.1 Flash hit 24/32 CVEs at pass@3 (75%) vs DeepSeek V4.1 Flash's 28/32 (87.5%), with 59.4% vs 70.8% single-run recall, ~40% more false positives, and $4.20 vs $2.40 per CVE; the author does not recommend it for security work.

The independent RSI-Inference Bench places Space Bunny second among open-weight models behind DeepSeek V4.1 Flash, and first among open models on its inference optimization track at 1.26x.

The post notes MiniMax M3.1 Flash is a 428B MoE that Unsloth's 2-bit dynamic GGUF shrinks to 143.9GB with 91.5% top-1 token agreement across seven domains, lowering the RAM barrier though not to laptop level.

Across 100 coding environments, M3.1 Flash lands at similar pricing to MiMo V2.6 Flash with lower performance, and the author says it is far from the Pareto front for open-weight or proprietary models.

An evaluator puts DeepSeek V4.1 Flash 1017 at 3.038, in a near-tie cluster with Meta Muse Wonder, Gemini 4 Argon, Mimo v2.6 Pro and MiniMax M3.1 Flash, strong at obstruction finding but weak on theorem scope.

The author ran a computer/browser-use benchmark in Runable and found MiniMax M3.1 Flash scored exactly the same as the Anthropic model — which even failed one task — yet the tool still ranked Anthropic higher, raising bias concerns.

A 4–6 hour complex test on a 600K+ word database reports success rates of 71% for GPT-6.1 Sol, 51% for MiniMax M3.1 Flash, 49% for Kimi K3.1 Pro, 39% for DeepSeek V4.1 Flash and 27% for GLM-5.3.

Citing Questflow's latest eval, M3.1 Flash rises from 73.8% to 94.9%, putting it on the same ~95% line as DeepSeek and Qwen Flash models, while noting real business bottlenecks are now dirty data and flaky APIs.

On Questflow's finance-agent benchmark covering retrieval, metric calculation and multi-tool calls, MiniMax M3.1 Flash rises from 73.8% to 94.9%, a few points behind Claude Opus 5.5 and free, with the author arguing the bottleneck has shifted to system-level data and tooling.

On 44 auto-graded time-limited agent tasks, DeepSeek V4.1 Flash scored 40/44, GLM-5.3 37/44, GLM-5.3-Flash 33/44, and MiniMax M3.1 Flash 26/44 but with the best vision; low effort beat max for DeepSeek and GLM.