Unsloth 推出 MiniMax M3.1 Flash 动态 GGUF：2-bit 版本仅 143.9GB

| NL2Repo-Bench | 61.2 | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 | | SWE Refactor Bench | 23.0 | — | 11.8 | 19.5 | 10.4 | 28.5 | 31.0 | | CyberGym | 82.2 | — | — | 80.0 | 84.5 | 83.6 | — | | BabyVision (with tools) | 91.0 | 92.5 | 89.6 | 85.7 | — | 88.9 | 94.1 | | HLE (without tools) | 43.4 | 42.6 | 39.2 | 43.5 | 42.3 | 49.5 | 55.6 | | CharXiv RQ | 82.5 | — | 69.0 | 84.8 | — | 85.2 | 83.7 | [...] | Benchmark | MiniMax M3.1 Flash | DeepSeek V4.1 Flash-1017 | DeepSeek V4.1 Flash | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 | --- --- --- --- | | Automation-Bench v1.0.6 | 62.8 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 | | BankerToolBench | 76.2 | — | 73.7 | 69.1 | 74.5 | 74.6 | 76.2 | | GDP.pdf AA | 20.8 | — | 12.8 | 22.0 | 11.2 | 27.2 | 21.6 | | Terminal Bench 4.0 | 38.1 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 | | LOCA-Bench 256K | 60.3 | — | — | 49.7 | 50.3 | 60.0 | 53.7 | | Agents’ Last Exam | 29.6 | — | — | 28.3 | — | 30.6 | 32.2 | | OSWorld 2.0 | 61.7 | — | 54.3 | 58.3 | — | 65.7

| Model | Total / active | Licence (points) | Index min. RAM | M3 Ultra | M5 Ultra | Score (rank) | --- --- --- | MiMo-V2.6-Flash | 309B / 15B | MIT (10) | 170 GB | community59 tok/s | community62 tok/s | 78 (#7) | | DeepSeek V4.1 Flash | 552B / 16B | MIT (10) | 434 GB | community20 tok/s | estimated~29 tok/s | 70 (#17) | | MiniMax-M3 | 428B / 23B | MiniMax Community (4) | 162 GB | community29 tok/s | estimated~42 tok/s | 65 (#44) | | Muse Spark Open | 816B / 42B | Muse Model License (5) | 342 GB | sourced18 tok/s | estimated~31 tok/s | 65 (#45) | | MiniMax-M3.1-Flash | 428B / 23B | MiniMax Community (4) | 189 GB | community25 tok/s | community35 tok/s | 64 (#53) |

Our 129.4GB 1-bit achieves ~89.9% top-1 accuracy; 143.9GB 2-bit reaches 91.5%, and 249.4 GB 4-bit UD-Q4\_K\_XL 98.0%. On coding, Q2 reaches 91.2% and Q4 97.6%. We weighted quant performance across seven categories: coding, math, chat, tools, multilingual, reasoning and agentic use. Thank you MiniMax for day-zero access. ### ⚙️ Usage Guide MiniMax-M3.1-Flash supports text, images and video and adds sparse attention across all layers, QAT for attention and routed experts, and DSpark. For 2-bit UD-Q2\_K\_XL, budget 160GB+ RAM + VRAM (or unified memory); for 4-bit UD-Q4\_K\_XL, budget 256GB+. These planning estimates assume 16K text context and one parallel slot. [...] Top-1 agreement by quant: Quant File size General chat top-1 Math top-1 `UD-IQ1_M` 129.4 GB 86.2% 94.9% `UD-IQ2_XXS` 134.9 GB 87.5% 95.3% `UD-Q2_K_XL` 143.9 GB 88.5% 95.7% `UD-IQ3_XXS` 160.0 GB 89.9% 96.3% `UD-IQ3_S` 175.5 GB 91.2% 96.8% `UD-Q3_K_XL` 194.0 GB 92.6% 97.4% `UD-Q4_K_XL` 249.4 GB

| 552B / 8B prefill, 16B decode (NVIDIA card) | 2.8T / 104B (HF) | ~744B / ~40B, shared GLM-5 base | | Context | 1M | 1M (DeepSeek docs) | 1,048,576 | 1,048,576 (Cloudflare) | | Inputs | Text, image, video | Text, image | Text, image | Text | | Weights / license | Due Oct 23; license not disclosed | Open, MIT | Open, Kimi K3 License | Open weights | | API price in / out (1M) | $0.15 / $0.60 flat | $0.30 / $1.20 peak; half off-peak | $3.00 / $15.00 | $1.40 / $4.40 | | Self-host hardware | Not disclosed | Not disclosed | MXFP4 weights; vLLM, SGLang | Not disclosed | | Automation-Bench v1.0.6\ | 62.2 | 54.8 | 46.7 | 48.8 | | DeepSWE v1.1\ | 73.7 | 74.2 | 67.5 | 66.9 | | Terminal Bench 4.0\ | 38.1 | 31.2 | 12.6 | 37.9 | | BabyVision (w/ tools)\ | 91.0 | 89.6 | 85.7 | Not reported | [...] The model accepts text, image and video input and returns text. The context window is 1,000,000 tokens. Those parameter totals match MiniMax M3, which ships at ~428B total and ~23B active. MiniMax has not

| 552B / 8B prefill, 16B decode (NVIDIA card) | 2.8T / 104B (HF) | ~744B / ~40B, shared GLM-5 base | | Context | 1M | 1M (DeepSeek docs) | 1,048,576 | 1,048,576 (Cloudflare) | | Inputs | Text, image, video | Text, image | Text, image | Text | | Weights / license | Due Oct 23; license not disclosed | Open, MIT | Open, Kimi K3 License | Open weights | | API price in / out (1M) | $0.15 / $0.60 flat | $0.30 / $1.20 peak; half off-peak | $3.00 / $15.00 | $1.40 / $4.40 | | Self-host hardware | Not disclosed | Not disclosed | MXFP4 weights; vLLM, SGLang | Not disclosed | | Automation-Bench v1.0.6\ | 62.2 | 54.8 | 46.7 | 48.8 | | DeepSWE v1.1\ | 73.7 | 74.2 | 67.5 | 66.9 | | Terminal Bench 4.0\ | 38.1 | 31.2 | 12.6 | 37.9 | | BabyVision (w/ tools)\ | 91.0 | 89.6 | 85.7 | Not reported | [...] Open Source Close Welcome Bar ### Get AI Dev Signals The newsletter 70,000+ AI builders read. Model releases, tools and research, without the hype. Exit mobile version

Title: MiniMax M3.1 Exposed: The 428B MoE Model That FORCES You to Think! # MiniMax M3.1 Exposed: The 428B MoE Model That FORCES You to Think! Think you know how LLM scaling works? MiniMax M3.1 Flash-Preview just flipped the script on open inference. By ditching raw parameter bloat for dynamic test-time compute governance, this system delivers an insane sub-300ms Time-To-First-Token (TTFT) across a massive 1,000,000-token context window! In this exclusive deep systems teardown, we get under the hood of MiniMax's newest powerhouse:. • The 428B MoE Monster: Why activating just 23B parameters (5.37%) beats brute-force compute. • MiniMax Sparse Attention (MSA): How O(N*B*d) scaling eliminates quadratic latency crashes. • The HTTP 400 Shock: Anthropic & OpenAI let you turn off thinking—MiniMax makes reasoning mandatory! • Triton Kernel Magic: 9x prefill speedup and 15x decoding acceleration in action. • Real-World Agent Chaos: 1,959 tool calls that optimized CUDA FP8 kernels from 7.6% to 71