# MiniMax M3.1 Flash — what changed, why it matters **Bottom line:** MiniMax shipped M3.1 Flash on **October 19, 2026** as a cheaper, much faster "Flash" tier of its M3 family, aimed squarely at agentic/office-work and coding tasks. It also revealed that the anonymous stealth model **"Space Bunny"** was an early build of it. The headline claims are ~3× faster inference and ~70% lower agentic-task cost — but nearly all benchmark numbers are MiniMax's own, and independent testing is only just starting. ## What changed **1. It's the same architecture, tuned for efficiency.** M3.1 Flash is a **428B-parameter MoE with ~23B active parameters**, 1M-token context, text/image/video input → text output — the same parameter counts as M3. MiniMax has **not published a technical report** on what specifically changed in the Flash variant ([MarkTechPost](https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/)). The efficiency gain comes from "more aggressive sparsity," which MiniMax says pushes output to **200+ TPS in everyday use, with P50 of 170+ TPS** ([MiniMax blog](https://www.minimax.io/blog/minimax-m31f)). **2. Pricing is flat and roughly half of M3.** $0.15/M input, $0.60/M output, $0.015/M cached input — with **no long-context surcharge**. M3 is $0.30/$1.20 up to 512K input and doubles above that ([MiniMax blog](https://www.minimax.io/blog/minimax-m31f); [MarkTechPost](https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/)). **3. Reasoning is always on and can't be disabled.** Five effort levels (Low → Max, with Max recommended as default). Requests sending `thinking.type=disabled` or `effort=none` now **return an error**, and reasoning arrives in a separate `reasoning_content` field — so integrations that parse inline `<think>` tags from M3 need updating ([MarkTechPost](https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/)). **4. The "Space Bunny" reveal.** MiniMax confirmed the anonymous model that ran on OpenRouter/OpenCode was an early M3.1 Flash build, processing **200T+ tokens** in stealth. OpenRouter says Space Bunny was its **#1 model for 10 straight days**, handling **56.7T tokens** — the most of any stealth model it has hosted. User feedback during that run drove measurable fixes: **TTFT 1.95s → 1.22s (−38%), throughput 70 → 100 TPS (+45%), error rate 1.4% → 0.7%** ([Unite.AI](https://www.unite.ai/minimax-launches-m3-1-flash-revealing-space-bunny-stealth-model/); [MiniMax blog](https://www.minimax.io/blog/minimax-m31f)). **5. Reported benchmark highlights** (all MiniMax-reported): Automation-Bench v1.0.6 **62.2** (vs Opus 5 at 50.3, GPT-5.6 Sol 45.8), BankerToolBench **76.2**, DeepSWE v1.1 **73.7%**, OSWorld V2 **61.7**, Agents' Last Exam **29.6**, CyberGym **82%**, BabyVision-with-tools **91.0** ([MiniMax blog](https://www.minimax.io/blog/minimax-m31f); [MarkTechPost](https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/)). ## Why it matters - **Cost-per-task, not cost-per-token, is the new battleground.** MiniMax frames value as dollars per completed task; on its Automation-Bench chart all five effort levels land under **$0.05/task**, and a full 152-task ALE run cost ~$111 ([MarkTechPost](https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/)). That's the metric that actually matters for agents that burn tokens on long tool-call chains. - **Independent testing is surprisingly favorable on intelligence.** Artificial Analysis scores M3.1 Flash **43 on its Intelligence Index — #4 of 173 models** — at **$0.22 cost per Intelligence Index task**, calling it "amongst the leading models in intelligence and well priced." Caveat: AA measures **123 TPS**, well below MiniMax's 170–200+ claim, and flags the model as **very verbose** (180M output tokens on the index vs. a 99M median) ([Artificial Analysis](https://artificialanalysis.ai/models/minimax-m3-1-flash)). - **It's a direct shot at DeepSeek.** The $0.15/$0.60 price is *exactly* DeepSeek V4.1 Flash 1017's off-peak rate. MiniMax's "flat rate" advantage is real mainly during DeepSeek's peak windows (6:30–9:30 AM and 11:30 AM–3:30 PM IST weekdays) — and MiniMax's own table shows DeepSeek winning several coding tests: DeepSWE 77.0 vs 73.7, Terminal Bench 4.0 42.3 vs 38.1, NL2Repo 68.2 vs 61.2 ([TechcityAuthority](https://www.techcityauthority.com/2026/10/minimax-m3-1-flash-price-performance.html)). - **Commercial stakes.** MiniMax (HK-listed since January) nearly quadrupled revenue to $116.6M in H1 2026, with **63.4% from API/enterprise** — while still posting a $293M adjusted net loss. Cheap Flash-tier inference is its growth engine ([TechcityAuthority](https://www.techcityauthority.com/2026/10/minimax-m3-1-flash-price-performance.html)). ## Caveats worth flagging - **Benchmarks are self-reported.** MarkTechPost notes "no independent replication exists yet," and MiniMax's own cost chart plots Automation-Bench at 62.8 while its table says 62.2. Terminal Bench 4.0 at **38.1** is its weakest result (Opus 5: 51.8) ([MarkTechPost](https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/)). - **Weights timing is inconsistent.** The blog says "later"; MiniMax's X post and MarkTechPost say **October 23, 2026**. The license terms are also unannounced — M3 used the MiniMax Community License (attribution required; written permission above $20M revenue), far stricter than DeepSeek's MIT ([Unite.AI](https://www.unite.ai/minimax-launches-m3-1-flash-revealing-space-bunny-stealth-model/); [TechcityAuthority](https://www.techcityauthority.com/2026/10/minimax-m3-1-flash-price-performance.html)). - **The 200T-token figure is unverifiable** — only OpenRouter's 56.7T slice is independently attributable ([TechcityAuthority](https://www.techcityauthority.com/2026/10/minimax-m3-1-flash-price-performance.html)). - **Availability:** MiniMax API, M Plan, and MiniMax Code desktop now; OpenRouter offered it free for 72 hours post-launch, then at list price. A more capable **Pro** model is promised later ([MiniMax blog](https://www.minimax.io/blog/minimax-m31f); [Unite.AI](https://www.unite.ai/minimax-launches-m3-1-flash-revealing-space-bunny-stealth-model/)).