今天新增minimax-m3.1IMAGE
M3.1 Flash 网络安全跑分：pass@3 找到 24/32 CVE，落后 DeepSeek V4.1 Flash
独立安全基准实测：M3.1 Flash pass@3 命中 24/32 CVE（75%），DeepSeek V4.1 Flash 28/32（87.5%）；单次召回 59.4% vs 70.8%，误报多约 40%，每个 CVE 成本 $4.20 vs $2.40，作者不建议用于安全任务。
pilvar (Philippe Dourassov)
@pilvar2226 小时前收录原推发布于 2026.10.27
pilvar (Philippe Dourassov) @pilvar222原推发布于 2026.10.27
We ran MiniMax M3.1 Flash on our cybersecurity benchmark. It's still a few months behind others for cybersecurity. - At pass@3, it found 24/32 CVEs (75.0%). DeepSeek V4.1 Flash 1017 found 28/32 (87.5%) - Single-run recall is 59.4%, vs 70.8% for DeepSeek - Precision is 80.5%, meaning it reports ~40% more false positives than DeepSeek - Cost: $4.20 per CVE found, vs $2.40 for DeepSeek V4.1 Flash 1017 It's honestly not that bad, but compared to the alternatives we have today, I would not recommend using it for cybersecurity tasks. 1/3 🧵
x.com/pilvar222/status/2115079609238561173采集于 2026.10.28 12:03