[{"data":1,"prerenderedAt":53},["Reactive",2],{"news-story-news-1792426066996-33oeb":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":14,"content_zh":19,"content_en":20,"external_links":21,"source":31,"category":38,"impact_score":39,"views_count":40,"likes_count":41,"related_entities":42,"tags":46,"image_url":52},"news-1792426066996-33oeb","MiniMax 正式发布 M3.1 Flash：100 万超长上下文、百万 Token 仅 0.15 美元，WebDev 竞技场大幅跃升 106 分","MiniMax Releases M3.1 Flash: 1M Context Window, $0.15/M Token Pricing, Surging +106 Points in WebDev Code Arena","MiniMax 正式推出新一代旗舰轻量大模型 M3.1 Flash（此前以 Space Bunny 代号霸榜全球竞技场 200T+ Token）。模型支持 100 万 Token 原生上下文，百万输入/输出 Token 定价低至 0.15 / 0.60 美元（缓存读取仅 0.015 美元），推理速度较 M3 提升 3 倍且成本降低三分之二。在 LMSYS Code Arena: WebDev 中斩获 1588 分跃升 106 分，位列开源模型第 12 名，定于 10 月 23 日正式开源权重。","MiniMax officially announced M3.1 Flash, its next-generation lightweight frontier model (previously dominating leaderboards as the secret model Space Bunny across 200T+ tokens). Featuring a native 1M context window, flat-rate API pricing of $0.15 / $0.60 per 1M input/output tokens ($0.015 cache reads), and 3x faster inference at 1/3 the cost of M3. It surged +106 points to 1588 on LMSYS Code Arena: WebDev (#12 among open models), with full weights set for open release on October 23.",[10,11,12,13],"MiniMax 正式发布新一代旗舰轻便大模型 M3.1 Flash，揭秘其正是此前霸榜竞技场超 200T Token 的神秘代号 Space Bunny","支持 100 万原生超长上下文，百万输入/输出仅需 0.15 / 0.60 美元（缓存读取 0.015 美元），推理速度提升 3 倍且成本降低三分之二","在 LMSYS Code Arena: WebDev 中斩获 1588 分跃升 106 分，位列开源模型第 12 名，游戏开发（+179 分）与参考设计（+155 分）大幅飞跃","Artificial Analysis 独立测评显示 Intelligence Index 达 42 分且 HLE 达 48%，完成基准测试消耗 Token 较竞品减少近半，权重将于 10 月 23 日全面开源",[15,16,17,18],"MiniMax officially launched M3.1 Flash, unveiling that it was the mystery model Space Bunny evaluated across 200T+ tokens","Features native 1M context with flat-rate API pricing of $0.15/$0.60 per 1M tokens ($0.015 cache reads), offering 3x speed at 1/3 cost of M3","Debuts on LMSYS Code Arena: WebDev at 1588 Elo (+106 pts over M3), ranking #12 among open models with major leaps in Gaming (+179) and Reference Design (+155)","Artificial Analysis benchmark confirms Intelligence Index of 42 and HLE of 48% at half the token consumption of rivals; weights release on Oct 23","### 核心背景与行业痛点\n在自主编程 Agent 与长上下文推理领域，开发者面临的核心矛盾始终是“高智能水平”与“爆炸性推理成本/高延迟”之间的失衡。以 SWE-bench、多轮前端开发和端侧协同为代表的复杂 Agent 流程，往往需要连续调用几十乃至上百次大模型，传统前沿闭源模型的昂贵 API 费用和高昂首字延迟（TTFT）成为智能体规模化落地的最大痛点。此前在各大排行榜上处理超 200 万亿 Token 的神秘模型“Space Bunny”，今日由 MiniMax 揭开底牌——全新旗舰轻便模型 `MiniMax M3.1 Flash` 正式上线。\n\n### 架构亮点与底层机制\n1. **原生 100 万超长上下文与极致 Token 经济学**：\n   - 支持高达 1M Token 的原生上下文窗口，支持长文本和高分辨率图像原生多模态输入；\n   - 彻底刷新 Agent 运行经济性：API 输入仅需 **$0.15 / 1M Token**，输出 **$0.60 / 1M Token**，缓存读取（Cache Read）更是降至极致的 **$0.015 / 1M Token**；\n   - 推理速度相比前代 MiniMax-M3 提升 3 倍，推理综合成本直接削减三分之二；\n2. **Space Bunny 成熟架构与开源计划**：\n   - 官方确认：该模型正是此前在 LMSYS 全球大模型匿名竞技场上测试超 200T Token、代号为“Space Bunny”的演进版本；\n   - 官方已正式宣布将于 10 月 23 日全面开放模型权重，成为具备 1M 上下文和顶级代码能力的重量级开源力量。\n\n### 权威 Benchmark 与实测跑分对比\nMiniMax 官方公布及 LMSYS Code Arena 和 Artificial Analysis 第三方独立实测数据显示（Figure 1, Arena WebDev, Artificial Analysis Benchmark）：\n- **全栈 Agent 核心能力评测（MiniMax 官方公布）**：\n  - **协同办公（Cowork）**：`Automation-Bench` 斩获 **62%**，`ALE` 达到 **30%**；\n  - **代码推理与工程修复（Coding）**：`DeepSWE` 达到 **74%**；\n  - **复杂多模态视觉推理（Visual Reasoning）**：`BabyVision` 达到 **91%**；\n- **LMSYS Code Arena: WebDev 实测对决**：\n  - 以 **1588 分**首次上榜，位列开源模型第 12 名（总榜第 32 名），助力 MiniMax 跃升全球前 12 大模型实验室；\n  - 较前代 MiniMax-M3（1482 分，排名第 19）实现 **+106 分**的巨大跨度；\n  - 在细分领域实现全方位跃升：**Gaming（游戏开发）** 排名从第 21 飙升至第 9（**+179 分**）；**Reference-Based Design（参考图设计）** 从第 20 跃升至第 7（**+155 分**）；**Simulations（环境模拟）** 跃升至第 13（**+122 分**）；**Content Creation Tools** 跃升至第 7（**+113 分**）；\n- **Artificial Analysis 独立中立基准评测**：\n  - **Intelligence Index（最高算力模式）** 达到 **42**（较 M3 大幅提升 13 点），追平 GLM-5.3-Flash（42），超越 Gemini 3.8 Flash（41），逼近 DeepSeek V4.1 Flash 1017（45）；\n  - **Humanity's Last Exam (HLE)** 达到 **48%**（超越 GLM 的 40% 与 DeepSeek 的 43%）；\n  - **CritPt** 达到 **22%**（大幅超越 M3 的 4% 与 GLM 的 15%）；\n  - **Token 效率与成本**：完成 Index 任务仅消耗约 1.89 亿 Token（仅为 Gemini 3.34 亿与 GLM 3.24 亿的近一半），单任务推理成本仅 **$0.22**（对比 Gemini 的 $1.24、M3 的 $0.51 与 GLM 的 $0.28）。\n\n### 开发者实战落地与开箱指南\n- **API 接入与平台**：开发者可直接访问 `platform.minimax.io` 开通 API，支持平滑兼容 OpenAI SDK 调用规范；\n- **桌面 Agent 原生支持**：MiniMax 同步推出了面向桌面开发者的专用客户端 MiniMax Code（`agent.minimax.io/download`），深度融合 M3.1 Flash 作为核心推理引擎；\n- **开源权重期待**：团队承诺 10 月 23 日在 Hugging Face 与开源社区同步释出开源模型权重，适合追求低延迟、高并发与低成本私有化部署的代码智能体团队。","### 核心背景与行业痛点\nIn autonomous coding agents and long-context reasoning pipelines, engineers constantly face a harsh tradeoff between frontier intelligence and prohibitive latency/token costs. Complex workflows on benchmarks like SWE-bench or multi-turn web application prototyping routinely execute dozens of iterative inference passes. High token tariffs and prolonged time-to-first-token (TTFT) from closed-source commercial APIs constrain production deployments. The mystery model \"Space Bunny\", which accumulated over 200 trillion tokens across public evaluation arenas, has been officially unveiled by MiniMax as `MiniMax M3.1 Flash`.\n\n### 架构亮点与底层机制\n1. **Native 1M Context Window with Unmatched Token Economics**:\n   - Supports up to 1,000,000 tokens of native context for both text and high-resolution visual inputs;\n   - Drastically resets agent operational economics: API pricing is set at **$0.15 / 1M input tokens** and **$0.60 / 1M output tokens**, with cache reads priced at an ultra-low **$0.015 / 1M tokens**;\n   - Yields 3x faster inference throughput compared to MiniMax-M3 while cutting overall execution costs by two-thirds;\n2. **Space Bunny Proven Lineage and Open-Weights Timeline**:\n   - Confirmed as the refined production release of \"Space Bunny\", battle-tested across hundreds of thousands of blind arena battles;\n   - MiniMax committed to releasing the open model weights on October 23, providing the open-source community with a flagship 1M-context model with top-tier code synthesis capabilities.\n\n### 权威 Benchmark 与实测跑分对比\nSynthesizing official benchmarks, LMSYS Code Arena, and independent Artificial Analysis evaluations:\n- **Core Agent Capabilities (Official Disclosures)**:\n  - **Cowork**: Achieves **62%** on `Automation-Bench` and **30%** on `ALE`;\n  - **Coding**: Reaches **74%** on `DeepSWE`;\n  - **Visual Reasoning**: Scores **91%** on `BabyVision`;\n- **LMSYS Code Arena: WebDev Blind Evaluation**:\n  - Debuts at **1588 Elo points**, securing #12 among all open models (#32 overall) and propelling MiniMax to the top 12 AI research labs worldwide;\n  - Represents a massive **+106 point increase** over MiniMax-M3 (1482 points, #19 open);\n  - Marked sub-category jumps include **Gaming** jumping from #21 to #9 (**+179 pts**), **Reference-Based Design** from #20 to #7 (**+155 pts**), **Simulations** to #13 (**+122 pts**), and **Content Creation Tools** to #7 (**+113 pts**);\n- **Artificial Analysis Independent Verification**:\n  - **Intelligence Index (Max Effort)**: Scores **42** (+13 over M3), matching GLM-5.3-Flash (42), beating Gemini 3.8 Flash (41), and trailing DeepSeek V4.1 Flash 1017 (45);\n  - **Humanity's Last Exam (HLE)**: Achieves **48%** (surpassing GLM's 40% and DeepSeek's 43%);\n  - **CritPt**: Scores **22%** (vs MiniMax-M3's 4% and GLM's 15%);\n  - **Token & Cost Efficiency**: Consumes only ~189M tokens across the full Index benchmark (roughly half of Gemini's 334M and GLM's 324M), delivering an industry-leading task cost of **$0.22** (compared to $1.24 for Gemini and $0.51 for M3).\n\n### 开发者实战落地与开箱指南\n- **API Availability**: Developers can obtain API keys at `platform.minimax.io`, supporting seamless drop-in compatibility with the OpenAI SDK specification;\n- **Desktop Agent Integration**: MiniMax released MiniMax Code (`agent.minimax.io/download`), a standalone desktop development assistant powered by M3.1 Flash;\n- **Open Weights Release**: Open weights will arrive on October 23 via Hugging Face for self-hosted deployments seeking low-latency agent infrastructure.",[22,25,28],{"title":23,"url":23,"type":24},"https://platform.minimax.io","other",{"title":26,"url":26,"type":27},"https://arena.lmsys.org","demo",{"title":29,"url":29,"type":30},"https://agent.minimax.io/download","docs",{"platform":32,"author_name":33,"author_handle":34,"author_avatar":35,"tweet_url":36,"published_at":37},"twitter","MiniMax","@MiniMax_AI","https://unavatar.io/x/MiniMax_AI","https://x.com/MiniMax_AI/status/2112150867868541221","2026-10-19T16:07:46.995Z","OpenSource",95,61,0,{"models":43,"agents":44,"vendors":45},[],[],[],[33,47,48,49,50,51],"M3.1 Flash","Space Bunny","Code Arena","WebDev","Artificial Analysis","https://pbs.twimg.com/media/HU_d5nbXEAAict2.jpg",1792905124113]