MiniMax-M3.1-Flash:From Anonymous Eval to Open Weights

Episode 1 · MiniMax-M3.1-Flash scores Intelligence Index 42, ranks 4th but shows 93% hallucination rate in AA evals (2026-10-19, 9 posts)

On Oct 19, Artificial Analysis published a full evaluation of MiniMax's new Flash reasoning model, MiniMax-M3.1-Flash, and the model was added to the Code Arena WebDev leaderboard. It scored 42 on the Intelligence Index—13 points above M3, on par with GLM-5.3-Flash and slightly above Gemini 3.8 Flash (41)—ranking 4th among 173 comparable models against a median of 13, at $0.15 per million input tokens ($0.22 per task). On Code Arena WebDev it scored 1588, placing 12th among open models and 32nd overall, up 106 points from its predecessor. The overall picture: first-tier intelligence at very low cost, but with notable reliability weaknesses.

Confirmed

• Intelligence Index 42, 13 points above M3, level with GLM-5.3-Flash, slightly above Gemini 3.8 Flash (41)
• 4th of 173 models (median 13); $0.15/M input tokens, $0.22 per task
• AA-Omniscience: 56% accuracy, ahead of GLM-5.3-Flash (28%) and DeepSeek V4 series, but a 93% hallucination rate
• Agentic results are split: AutomationBench-AA 73% leads GLM-5.3-Flash (max, 60%) and others, while terminal and real-world tasks still lag
• Artificial Analysis released all 10 Intelligence Index sub-evaluations (AA-Briefcase, GDPval-AA, Terminal, etc.)
• Code Arena WebDev: 12th open-source, 32nd overall, 1588 points, +106 vs previous generation
• User tests: decoding speed 68.8–105.7 tok/s, averaging near 100 tok/s, no provider errors; rated slightly slower than DeepSeek-v4.1-flash

Why it matters

• Ultra-low pricing plus first-tier intelligence makes it a strong candidate for cost-sensitive workloads, with meaningful per-task savings at scale
• The coexistence of a 93% hallucination rate with leading automation benchmarks suggests users should evaluate by task type: strong on automation, risky where factual reliability matters

• MiniMax-M3.1-Flash hits ~100 tok/s in speed tests, trailing DeepSeek v4.1-flash — TheZachMueller · 2026-10-19
• Code Arena WebDev: MiniMax-M3.1-Flash ranks #12 among open models, DeepSeek-V4.1-Flash tops at 1664 — arena · 2026-10-19
• MiniMax-M3.1-Flash debuts at #12 among open models in Code Arena with 1588 pts — cedric_chee · 2026-10-19
• MiniMax-M3.1-Flash scores 42 on AA Intelligence Index, halving cost per task to $0.22 — ArtificialAnlys · 2026-10-19
• MiniMax-M3.1-Flash scores 42 on Intelligence Index at ~$0.22 per task — ArtificialAnlys · 2026-10-19
• MiniMax-M3.1-Flash leads AutomationBench at 73% but trails on Terminal-Bench — ArtificialAnlys · 2026-10-19
• MiniMax-M3.1-Flash tops knowledge accuracy but hallucinates on 93% of unknowns — ArtificialAnlys · 2026-10-19
• Full per-evaluation breakdown for MiniMax-M3.1-Flash across the AA Intelligence Index — ArtificialAnlys · 2026-10-19
• MiniMax-M3.1-Flash ranks #4 on Artificial Analysis Intelligence Index at $0.15/M input tokens — ArtificialAnlys · 2026-10-19

Episode 2 · MiniMax Open-Sources M3.1-Flash: 428B Multimodal MoE with 1M Context (2026-10-23, 8 posts)

On October 23, MiniMax officially open-sourced MiniMax-M3.1-Flash and released its weights. The company confirmed that the model previously pre-trained anonymously under the codename "Space Bunny"—consuming 200T+ tokens—was in fact M3.1-Flash, an upgraded version of MiniMax-M3.

Confirmed

• The model is a natively multimodal MoE architecture with roughly 428B total parameters and 23B active parameters per token.
• It supports a 1 million-token context with sparse attention used throughout (the post mentions a Block-sparse style mechanism).
• It is positioned for coding, agentic workflows, and computer use; MiniMax says it excels at agentic tasks and safety.
• Weights are available on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash) and ModelScope.
• Ecosystem support: Together AI is a day-0 launch partner offering fully managed APIs; NVIDIA NeMo supports fine-tuning on up to 128 GPUs.

Not Yet Confirmed

• Reddit user @mrtime777 noted that the post lacks benchmark data and details; the model's actual performance and pricing remain to be announced.

Why It Matters

This is another large open-source multimodal MoE model, pre-trained anonymously (as "Space Bunny") on 200T+ tokens. The combination of ultra-long context, native sparse attention, and an Agent/computer use focus shows open-source models are now targeting frontier use cases like tool calling and automated workflows, while day-0 ecosystem support from Together and NVIDIA lowers the barrier to adoption.

• MiniMax Quietly Lists New M3.1-Flash Model on Hugging Face — mrtime777 · 2026-10-23
• MiniMax-M3.1-Flash goes open weights with day-0 NVIDIA NeMo fine-tuning up to 128 GPUs — MiniMax_AI · 2026-10-23
• MiniMax open-sources M3.1 Flash: 428B total params, 23B active, built for agentic tasks — VoidAsuka · 2026-10-23
• MiniMax ships M3.1 Flash: native multimodal MoE with 1M context and sparse attention throughout — KyeGomezB · 2026-10-23
• MiniMax open-sources M3.1 Flash: 428B MoE with 23B active, 1M context, live on Together AI — MiniMax_AI · 2026-10-23
• MiniMax Open-Sources M3.1 Flash: 428B Params, 23B Active, NVIDIA Vera Rubin Day-0 Support — NVIDIAAI · 2026-10-23
• MiniMax M3.1 Flash ships with day-0 vLLM support and NVFP4 serving stack — woosuk_k · 2026-10-23
• MiniMax site lists wave of new releases: M3.1 Flash, M3, video model H3, Speech 2.8 and more — lipeng0820 · 2026-10-23

Episode 3 · vLLM Announces Day-0 Support for MiniMax M3.1 Flash with NVFP4 Gains (2026-10-23, 4 posts)

vLLM provides day-0 serving support for MiniMax M3.1 Flash, with native NVFP4 cutting checkpoint size by 44.6% and boosting 256K-context concurrency 2.7x, reaching 191.5k tok/s per GPU on Vera Rubin NVL72 alongside strong benchmark results.

• M3.1 Flash hits 191.5k tok/s per GPU on Vera Rubin NVL72, leads open models on agent benchmarks — vllm_project · 2026-10-23
• MiniMax M3.1 Flash benchmarks: 91.0 BabyVision, 48% HLE, top BankerToolBench score — vllm_project · 2026-10-23
• vLLM details M3.1 Flash serving: NVFP4 cuts weights 44.6%, boosts capacity 2.7x — vllm_project · 2026-10-23
• vLLM ships Day-0 serving for MiniMax M3.1 Flash: NVFP4 cuts KV memory 43.75%, 2.7× concurrency — vllm_project · 2026-10-23