[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1zdqrec1j7fzn":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":24,"bodyZh":25,"bodyEn":26,"sourceName":27,"sourceUrl":28,"updatedAt":29},"minimax-m31-flash-2026-10-20","NEWS","2026-10-20T10:00:00.000Z","MiniMax-M3.1-Flash 发布：API 先行、权重后置，在线 3 倍速、成本降 70%","MiniMax-M3.1-Flash ships: API first, weights later, 3x faster online at 70% lower cost","10 月 19 日 MiniMax 上线 M3.1-Flash（Flash 系列新旗舰，API 先行）。官方口径：在线推理约 3 倍速、任务成本降 70%（$0.15 \u002F $0.60 \u002F $0.015 每百万 token），输出 200+ tok\u002Fs。匿名声测代号 Space Bunny，已处理 200 万亿 token。权重稍后放出——站内按 RESTRICTED \u002F 参数未公布口径收录，零实测记录。","MiniMax shipped M3.1-Flash on Oct 19, the new Flash-series flagship, API first. Official claims: roughly 3x faster online inference, 70% lower task cost ($0.15 \u002F $0.60 \u002F $0.015 per 1M tokens), 200+ tok\u002Fs output. It ran anonymously as Space Bunny, processing 200 trillion tokens. Weights come later — logged as RESTRICTED with unpublished params, zero on-site runs.",[12,15,18,21],{"id":13,"name":14},9,"GLM-5.3-Flash",{"id":16,"name":17},11,"Qwen3.8-Flash-Next",{"id":19,"name":20},22,"MiniMax-M3",{"id":22,"name":23},91,"MiniMax-M3.1-Flash",[],"# MiniMax-M3.1-Flash 发布：API 先行、权重后置，在线 3 倍速、成本降 70%\n\n10 月 19 日，MiniMax 上线 **M3.1-Flash**——Flash 系列新旗舰，走的是「API 先行、权重后置」路线：今天就能调，本地要等官方放出权重。\n\n## 速度与价格（官方口径）\n\n| 指标 | M3.1-Flash | 相对前代 |\n|---|---|---|\n| 在线推理速度 | ~3 倍 | 约 3x |\n| 任务总成本 | — | 降 70% |\n| 输出吞吐 | 200+ tok\u002Fs（P50 170+） | — |\n| 输入价格 | $0.15 \u002F 1M token | — |\n| 输出价格 | $0.60 \u002F 1M token | — |\n| 缓存输入 | $0.015 \u002F 1M token | — |\n\n更激进的稀疏化是这次提速的来源：同任务用更少激活参数完成，官方同时把「成本降 70%」作为核心卖点——对高频调用的 agentic 场景，这比单看基准分更实在。\n\n## 基准（官方口径）\n\n| 基准 | 得分 |\n|---|---|\n| BankerToolBench | **76.2** |\n| DeepSWE v1.1 | 73.7% |\n| OSWorld V2 | 61.7 |\n| Agents Last Exam | **29.6**（中国模型双第一） |\n| CyberGym | 82% |\n\n值得注意的是 Agents Last Exam 双榜（总体 + 中国模型）第一与 CyberGym 82%——Flash 档模型在安全 \u002F 漏洞类基准上拿到这个分数，说明稀疏化没有明显牺牲 agentic 长尾能力。\n\n## Space Bunny：先跑量、后揭幕\n\n发布前该模型以匿名声测代号 **Space Bunny** 在开放环境运行，已累计处理 **200 万亿 token**——官方用真实流量先验证，再揭晓名字。这套「先跑量后发布」的节奏最近在多家实验室流行，对用户的实际意义是：API 的稳定性在揭幕前已被压测过。\n\n## 站内收录口径\n\n站内已收录 **MiniMax-M3.1-Flash**（id 91）：参数量未公布 → `params_b = NULL`，权重未放出 → `RESTRICTED`。与 Kimi-K3 的 API-only 口径一致——**零实测记录**，本地速度、显存账等权重发布后再补。\n\nFlash 场坐标（站内已有目录行）：\n\n| 模型 | 总参 \u002F 激活 | 备注 |\n|---|---|---|\n| MiniMax-M3 | 427B | 前代旗舰 |\n| GLM-5.3-Flash | 320B \u002F A32B | 智谱 Flash 档 |\n| Qwen3.8-Flash-Next | 180B \u002F A14B | 阿里 Flash 档 |\n\n## 获取方式\n\n当前可用渠道：**MiniMax API \u002F M Plan \u002F MiniMax Code**。想本地部署的用户需要等权重放出；按 MiniMax 近几代的节奏，权重窗口通常在 API 发布后数周内。","# MiniMax-M3.1-Flash ships: API first, weights later, 3x faster online at 70% lower cost\n\nOn Oct 19, MiniMax shipped **M3.1-Flash**, the new Flash-series flagship, following an API-first, weights-later path: callable today, local runs wait on the official weights release.\n\n## Speed and price (official numbers)\n\n| Metric | M3.1-Flash | vs predecessor |\n|---|---|---|\n| Online inference speed | ~3x | ~3x |\n| Task total cost | — | down 70% |\n| Output throughput | 200+ tok\u002Fs (P50 170+) | — |\n| Input price | $0.15 \u002F 1M tokens | — |\n| Output price | $0.60 \u002F 1M tokens | — |\n| Cached input | $0.015 \u002F 1M tokens | — |\n\nMore aggressive sparsity drives the speedup: the same tasks finish with fewer active parameters, and the official pitch pairs that with a 70% cut in task cost — for high-frequency agentic workloads that matters more than a benchmark delta.\n\n## Benchmarks (official numbers)\n\n| Benchmark | Score |\n|---|---|\n| BankerToolBench | **76.2** |\n| DeepSWE v1.1 | 73.7% |\n| OSWorld V2 | 61.7 |\n| Agents Last Exam | **29.6** (first among Chinese models on both leaderboards) |\n| CyberGym | 82% |\n\nFirst place overall and among Chinese models on Agents Last Exam, plus 82% on CyberGym, show the sparsity did not visibly cost agentic long-tail capability.\n\n## Space Bunny: run first, reveal later\n\nBefore launch the model ran anonymously as **Space Bunny** in open beta, processing **200 trillion tokens** — the lab validated on real traffic before revealing the name. The run-first-reveal-later pattern is in vogue; for users it means the API stability was already load-tested before the banner went up.\n\n## How we logged it\n\nOur catalog now lists **MiniMax-M3.1-Flash** (id 91): parameter count unpublished → `params_b = NULL`, weights not out → `RESTRICTED`. Same API-only stance as Kimi-K3 — **zero on-site records**; local speed and VRAM accounting wait for the weights.\n\nFlash-field coordinates (existing catalog rows):\n\n| Model | Total \u002F active | Note |\n|---|---|---|\n| MiniMax-M3 | 427B | predecessor flagship |\n| GLM-5.3-Flash | 320B \u002F A32B | Zhipu Flash tier |\n| Qwen3.8-Flash-Next | 180B \u002F A14B | Alibaba Flash tier |\n\n## Availability\n\nAvailable now through the **MiniMax API, M Plan and MiniMax Code**. Local deployers wait on the weights; on MiniMax's recent cadence the weights window has usually opened within weeks of the API launch.","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fminimax-m31-flash-2026-10-20","2026-10-20T12:13:07.902Z"]