[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-minimax-m3-1-flash-space-bunny":3,"topics-all":38,"news-related-8c6d408b-2913-48b2-bc1f-4e958ad854c5":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"8c6d408b-2913-48b2-bc1f-4e958ad854c5","Space Bunny 谜底揭晓:MiniMax M3.1 Flash,匿名期吞下 200 万亿 token","MiniMax 发布 M3.1 Flash,揭晓匿名霸榜 OpenCode 12 天的 Space Bunny 即其早期版本,匿名期处理超 200 万亿 token。自报 DeepSWE 73.7%、OSWorld V2 61.7 分,agent 成本降 70%、输出 200+ TPS;权重稍后放出,尚无独立复现。","10 月 19 日,MiniMax 官宣 M3.1 Flash,顺手揭晓了一个悬念:此前以 Space Bunny Alpha 之名在 OpenRouter 匿名上线、并连续 12 天占据 OpenCode 榜首的神秘模型,就是它的早期版本——匿名运营期间处理了超过 200 万亿 token。这不只是一次发布:模型先不带厂牌下场,拿真实负载证明自己,再亮明身份,等于把发布会的说服成本转嫁给了 200 万亿 token 的匿名用量。\n\n## 数字面:agentic 四线自报成绩\n\n按官方博客口径(以下均为厂商自报),M3.1 Flash 在四条 agent 赛道交卷:软件工程基准 DeepSWE v1.1 拿下 73.7%;电脑操作基准 OSWorld V2 61.7 分、Agents' Last Exam(ALE)29.6 分,官方称这两项在中国模型中排名第一;网络安全基准 CyberGym 82%;金融工具基准 BankerToolBench 76.2。效率侧,官方称靠「更激进的稀疏化」把在线推理提速近 3 倍,输出速度 200+ TPS、P50 稳定在 170+ TPS;第三方 CellCog 转述发布帖还标注了 100 万 token 上下文。\n\n## 价格与开放度\n\n定价为每百万 token 输入 0.15 美元、输出 0.60 美元、缓存命中 0.015 美元,官方称 agent 任务整体使用成本较上代降 70%。目前只能通过 MiniMax API、M Plan 与 MiniMax Code 使用;官方博客对权重只写了「稍后放出」,CellCog 读到的发布帖标注 10 月 23 日——兑现之前,「开源」仍是待兑现承诺。\n\n## 泼冷水:自报跑分与真实手感\n\n两个独立信号值得记下。其一,CellCog 整理的横向对比表里,M3.1 Flash 并非全面领先:DeepSWE v1.1 上 DeepSeek V4.1 Flash 报 77.0、Opus 5 报 74.0,都高于它的 73.7;Terminal Bench 4.0 它也明显落后 Opus 5(38.1 对 51.8)。其二,首批实测用户反馈偏冷:有开发者在 X 上表示在 OpenCode 写日常代码「感觉不如 GLM-5.3 Flash」,另一位用户反映它在工具调用长链路上会卡住,「单 token 便宜,但会话在第 40 步挂掉就没意义」。而 OpenCode 官方账号披露,Space Bunny 曾连续 12 天居其榜首、服务了 106 万亿 token。\n\n## 所以呢\n\nSpace Bunny 这套「匿名先行、亮牌收尾」的打法,聪明之处在于把跑分的公信力问题换成了一道更难的题:200 万亿 token 的真实用量很难造假。但自报 benchmark 尚无人独立重跑,匿名端点与正式 API 的表现是否一致也待验证;等权重(若如期在 10 月 23 日)落地,自己跑一遍 DeepSWE,才算真正交卷。\n\n参考:MiniMax 官方博客 https:\u002F\u002Fwww.minimax.io\u002Fblog\u002Fminimax-m31f ;CellCog 分析 https:\u002F\u002Fcellcog.ai\u002Fblog\u002Fminimax-m3-1-flash","https:\u002F\u002Fwww.minimax.io\u002Fblog\u002Fminimax-m31f","70524a06-fc44-487c-ac6b-4a0186f66a45",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"42fc8bd8-25ab-4a04-9678-34ac4b60396e","en","Space Bunny unmasked: MiniMax M3.1 Flash, 200T tokens","MiniMax unmasks Space Bunny as M3.1 Flash: 12 days atop OpenCode, 200T tokens. DeepSWE 73.7% and 70% cost cut are self-reported; weights later.","On October 19, MiniMax officially rolled out M3.1 Flash — and unmasked a mystery in the process: Space Bunny Alpha, the anonymous model that ran on OpenRouter and topped OpenCode's chart for 12 consecutive days, was an early version of it, processing more than 200 trillion tokens while unnamed. That is the real story: the model proved itself on real workloads before claiming its brand.\n\n## The numbers, all vendor-reported\n\nPer the official blog, M3.1 Flash posts 73.7% on DeepSWE v1.1 (software engineering), 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam (MiniMax says both rank first among Chinese models), 82% on CyberGym, and 76.2 on BankerToolBench. Inference is claimed to be nearly 3x faster online via \"more aggressive sparsity,\" with 200+ TPS output and a P50 of 170+ TPS. CellCog's read of the release post also notes a 1M-token context.\n\n## Pricing and openness\n\nFlat pricing is $0.15 per million input tokens, $0.60 output, and $0.015 cached; MiniMax claims a 70% drop in overall agent-task cost versus the predecessor. Access is via the MiniMax API, M Plan, and MiniMax Code. The blog only says weights arrive \"later\"; CellCog reads October 23 from the release post. Until then, \"open\" remains a promise.\n\n## The cold water\n\nTwo independent signals. First, CellCog's comparison table shows M3.1 Flash is not uniformly ahead: on DeepSWE v1.1, DeepSeek V4.1 Flash reports 77.0 and Opus 5 reports 74.0, both above its 73.7; on Terminal Bench 4.0 it trails Opus 5 (38.1 vs 51.8). Second, early hands-on feedback is lukewarm: one developer wrote on X that for everyday coding in OpenCode it \"still feels weaker than GLM-5.3 Flash,\" and another reported it stalling on long tool-call chains — \"cheap per token means nothing if the session dies at step 40.\" OpenCode's own account disclosed Space Bunny held its #1 spot for 12 days, serving 106 trillion tokens.\n\n## So what\n\nThe anonymous-first, reveal-later playbook swaps benchmark credibility for something harder to fake: 200 trillion tokens of real usage. But the vendor-reported scores have not been independently rerun, and whether the anonymous endpoint matches the official API is unverified. When the weights land — October 23, if on schedule — run DeepSWE yourself; that is the real acceptance test.\n\nRefs: MiniMax blog https:\u002F\u002Fwww.minimax.io\u002Fblog\u002Fminimax-m31f ; CellCog analysis https:\u002F\u002Fcellcog.ai\u002Fblog\u002Fminimax-m3-1-flash","minimax-m3-1-flash-space-bunny","2026-10-19T21:10:12Z","2026-10-19T21:10:44.876195Z","2026-10-19T21:10:44.876206Z",true,"agent",112,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a9191131-1fe3-42e4-870d-2621ef5874d1","5月LLM发布大爆炸：300款模型混战，重塑AI产业格局","may-2026-300-llm-release-explosion","2026-05-28T08:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"8c27ba0d-9f70-4e4c-9a9b-092eec0db248","Grok 4.3 登陆 OCI：xAI 与 Oracle 的「次日达」合作意味着什么","grok-4-3-oci-oracle-next-day","2026-05-15T04:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"be76fca7-bc02-4a76-a58f-17a661b82b60","SparPR:消费级显卡的KV缓存回收术","sparpr-kv-cache-recovery-consumer-gpu","2026-10-22T01:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"8c579d8d-4114-4d7d-92c1-22f5fab0537b","连续流语言模型跑到 8B：Lafa 单模型自验证解码提速 2.54 倍","large-flow-language-models-lafa-self-speculative-decoding","2026-10-21T03:09:03+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"713ebd28-7abc-4664-acd5-8ba7c6b50d97","扩散 LLM 能否 scale?Mercury 3 用 40 倍规模和 1M 上下文作答","mercury-3-diffusion-scaling","2026-10-19T23:09:33+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"559df680-f13e-48f0-9263-73116aaff411","上交大DLite免历史投机解码:Agent提速2.74倍","dlite-history-free-speculative-decoding","2026-10-19T17:07:52+00:00"]