MiniMax launches M3.1 Flash: 3x faster at 1/3 the cost, weights open-sourcing Oct 23

MiniMax officially released M3.1 Flash on October 19, positioning it as "frontier intelligence at extreme value": one-third the cost of M3, 3x faster, with a flat API price of $0.15, a 1M-token context window, adjustable thinking depth, and model weights set to be open-sourced on October 23. The company confirmed that the mysterious model "Space Bunny"—which topped multiple global leaderboards and processed over 200T tokens—was in fact an early version of M3.1 Flash.

Confirmed

• Official benchmark results: Cowork Automation-Bench 62% (per m6, beating Opus 5's 50.3% and GPT-5.6 Sol's 45.8%), DeepSWE coding 74%, visual reasoning BabyVision (with tools) 91.0, CharXiv RQ 82.5, HLE 48%, AutomationBench-AA 72.5% (top among open-source models in launch comparisons)
• vLLM offers Day-0 support and disclosed serving-stack optimizations: MiniMax Sparse Attention (MSA) natively uses NVFP4 attention, with sparse attention delivering up to 1.7x speedups; in partnership with Inferact, Indexer DCP shards context across multiple GPUs, speeding up 512K-context serving by 3.3x and cutting TTFT by a further 19%
• testingcatalog reports MiniMax also launched H3, a new-generation open-source general multimodal video model, around the same time

Unconfirmed

• teortaxesTex's analysis suggests M3.1 Flash may be even smaller than V4.1, and notes DeepSeek V4.1F's active parameter count hit a series low, speculating that compute scarcity is shrinking model sizes—these benchmarks and analysis are not officially verified

Why it matters

With flat $0.15 pricing and an open-source commitment, M3.1 Flash is taking aim at the cost-effective agentic model market, with multi-step automation benchmark scores reportedly overtaking closed-source flagships. If the industry thesis that "compute shortages are forcing models to slim down" holds, small-but-mighty could become the dominant theme of the next phase of model competition.

2026-10-19 ~ 2026-10-20 · 25 related posts

Full story(2 episodes)→

• Episode 1: MiniMax Teases Return of "Space Bunny" with Upcoming M3.1-Flash Model(2026-10-19, 3 posts)
• Episode 2: MiniMax launches M3.1 Flash: 3x faster at 1/3 the cost, weights open-sourcing Oct 23(2026-10-19, 25 posts)

• MiniMax Launches Open Multimodal Video Model H3 and M3.1-Flash With 1M Context — testingcatalog · 2026-10-19
• MiniMax Launches M3.1 Flash: 3x Faster Inference, 70% Cheaper, Top Agentic Benchmarks — MiniMax 稀宇科技 · 2026-10-19
• [source] MiniMax launches M3.1 Flash: 3x faster, 1/3 the cost of M3, weights out Oct 23 — MiniMax_AI · 2026-10-19
• MiniMax M3.1 Flash debuts at 1/3 the cost of M3 with 3x speed, weights land Oct 23 — MagicZhang · 2026-10-19
• MiniMax M3.1 Flash leaks smaller than V4.1 as compute scarcity pushes models downsize — teortaxesTex · 2026-10-19
• DeepSeek's V4.1F has fewer active params than any prior flagship — compute scarcity biting — teortaxesTex · 2026-10-19
• MiniMax M3.1 Flash targets multi-step work: 72.5% AutomationBench-AA, best open model — vllm_project · 2026-10-19
• MiniMax M3.1 Flash benchmarks: 91.0 BabyVision, 48% HLE, 76.15 BankerToolBench — vllm_project · 2026-10-19
• [source] vLLM details MiniMax M3.1 Flash serving stack: 1.7x faster indexer decode via NVFP4 — vllm_project · 2026-10-19
• vLLM + MiniMax: 3.3x faster 512K-context indexer serving, 19% lower TTFT on GB300 — vllm_project · 2026-10-19
• China's Flash twins clash: MiniMax M3.1 Flash lands two days after DeepSeek's — 新智元 · 2026-10-19
• MiniMax M3.1 Flash launches, tops AutomationBench among open-weight models — Aizkmusic · 2026-10-20
• MiniMax ships 428B MoE model M3.1 Flash at $0.15 per million input tokens — VectorWireAI · 2026-10-20
• MiniMax M3.1 Flash goes live on Venice, 3x faster than M3 with anonymous access — Scobleizer · 2026-10-20
• MiniMax M3.1 Flash benchmark table leaks: tops automation evals at 62.8, size undisclosed — QuixiAI · 2026-10-20
• MiniMax Launches M3.1 Flash Without Disclosing Model Size, Users Ask — QuixiAI · 2026-10-20

9 near-duplicate retellings: kimmonismus · MiniMax_AI · haider1 · MiniMax_AI · APPSO · AccBalanced · hey_abusiddik · FlightUsed648 · woosuk_k