[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-thinkingcap-gemma-4-26b-shorter-thinking":3,"topics-all":38,"news-related-0ecc4f65-2a46-4e7f-8694-510db4f299cf":67},{"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},"0ecc4f65-2a46-4e7f-8694-510db4f299cf","ThinkingCap:Gemma-4 少想 38.9%","ThinkingCap 在 Google 的 gemma-4-26B-A4B-it 上做后训练,把思考 token 平均砍掉 38.9%,十二项基准平均准确率只掉 0.15pp。官方另给出另一款模型在单卡上 4.7 倍吞吐的实测,以及长上下文掉 4pp 的代价。","12 个基准跑下来,准确率只掉了 0.15 个百分点,思考 token 却少了 38.9%。BottleCap AI 把这版后训练模型叫 ThinkingCap,底座是 Google 的 gemma-4-26B-A4B-it,权重和 GGUF 量化版都放到了 Hugging Face,换上就能跑([原始公告](https:\u002F\u002Fbottlecapai.com\u002Fpost\u002Fthinkingcap-gemma-4-26b\u002F))。\n\nThinkingCap 系列的第五个模型,也是第一个落在 Google 模型上的——之前四款分别在 Qwen3.6-27B、Qwen3.8-27B、NVIDIA Nemotron-3.5-Lightning-30B-A3B 和 Qwen3.8-Flash-Next 上做过。\n\n## 数字出自谁的手\n\n这是厂商自己的评测,BottleCap 在公告里说得很直白:基座模型不是引用公开数据,而是他们自己在同一套评测里跑出来的,采样参数、温度、top_p、top_k 全部沿用 Google 的推荐设置。换句话说两侧是同一场比赛,不是拿自家模型去比别家论文里的数字。\n\n测法也交代了:AIME 2026、HMMT 两套、LiveCodeBench v6、GPQA-Diamond、MMLU-Pro、MMMLU、AA-LCR、LongBench v2、RealWorldQA、IFBench、tau2-bench,一共 12 个;每个基准跑的随机种子从 4 个到 64 个不等,准确率报 95% 置信区间;思考写不完直接没答案的\"截断\",以及模型反复兜圈不收尾的\"循环\",也单独统计。\n\n在这个口径下,12 项里有 8 项的差距小于运行本身的波动范围,5 项甚至反超基座。真正跑出置信区间的只有 4 项,方向还不一致:IFBench 反而涨了 1.83 个百分点,GPQA-Diamond 掉 0.96,MMMLU 掉 0.66,AA-LCR 掉 4.00。最后这一项是整套数字里唯一的实打实的代价。\n\n## 省下来的时间去了哪\n\n平均砍 38.9% 不是均匀的。长上下文和多模态那组砍得最狠,54.4%;知识和推理 44.4%;数学 36.1%;代码 27.6%;指令遵循和 agentic 只砍 17.2%。单项最猛的是 LongBench v2 的 69.1%,最浅的 IFBench 只有 13.7%。\n\n值得单独拎出来的是两个失败率:思考撞上 65536 token 生成上限、直接没答案的截断,从 4.35% 降到 0.72%;反复兜圈不收尾的循环,从 3.64% 降到 0.76%。对跑生产环境的人来说,这两个数字可能比 38.9% 更实在。\n\n还有一处细节值得警惕:掉得最狠的 AA-LCR,恰好也是思考被砍得最狠的那一组。长上下文少想了 59.3%,准确率就掉 4.00 个百分点。也就是说这版模型省下的思考,主要是从长上下文能力里抠出来的。要是你正要上一套重度依赖长文检索的 Agent,这个交换未必划算。\n\n## 落到账单上,关键是 KV cache\n\n按 token 单价算,少 38.9% 只是一半账。推理模型部署时真正卡住服务器的常常不是算力,而是 KV cache:每条并发序列都要占一份缓存,缓存装不下,能同时跑的任务数就被卡死。\n\nBottleCap 另一篇文章做了吞吐实测:同一张 RTX PRO 6000、8GB KV cache、vLLM 0.29.0、关掉前缀缓存、64 并发、20 分钟窗口,同一个 AIME 几何题,原版 Qwen3.8-27B 跑完 24 个任务,ThinkingCap 版跑完 112 个,4.7 倍。思考长度只有原来的 2.3 分之一,换来的是平均并发从 15.1 抬到 24.9、token 吞吐从 319 到 511 tok\u002Fs。机制不复杂:trace 短了,单条序列占的缓存就少,同一张卡塞进更多并发,每条又更早结束,缓存周转更快。\n\n需要说明的是,这个 4.7 倍是在另一个底座(Qwen3.8-27B)上量的,不是 Gemma-4。但 Gemma-4 底座本身更值得优化:第三方评测 CrucibleMark 在 DGX Spark 上测 Think 模式,综合分 74.94%,P95 响应 91.09 秒,49 题里有 1 题超时,评价是偏\"批量 DevOps 专家\"型的选手——慢但有料,正好是吃 KV cache 的典型负载。按第三方记录,该模型是 MoE,总参 25.2B、每 token 激活约 3.8B,Apache 2.0 授权,上下文 262,144,Hugging Face 上也已有现成的 GGUF 量化版。\n\n## 谁该现在换,谁该等等\n\n如果你的服务是长文档问答、代码库检索、深度研究这类长上下文活,先别急着换,AA-LCR 那 4 个百分点会先告诉你疼在哪。如果负载是批量跑、后台任务,或者 agent 循环,那截断和循环率掉一个数量级,可能比准确率上那 0.15 个百分点更值钱。\n\n两个现实约束:一是这版是第三方后训练产物,底座 Apache 2.0,但衍生权重的许可条款仍以 Hugging Face 上那份 model card 为准,商用前自己看清楚;二是宣称\"最多再省 87.6% 思考 token\"的 PRO 版本只走商务邮箱,没有公开权重。\n\n我更在意的其实是第三点:所有 ThinkingCap 数字目前都出自厂商自己的评测框架,社区还没有第三方复现。而\"省 token 又不掉准确率\"这类宣称,历来是最需要独立验证的一类。所以我的建议是,先在真实流量上跑一遍 A\u002FB,别把 38.9% 当成既定事实。\n\n所以呢?推理模型的成本优化,从\"让模型少想点\"走向\"让同一张卡多干点活\",这才是真正该被讨论的方向——而这家公司已经用两篇公告,把这两件事拆开摆给你看了。\n","https:\u002F\u002Fbottlecapai.com\u002Fpost\u002Fthinkingcap-gemma-4-26b\u002F","654deb1c-1bb4-43cf-8907-ec8c4ae65a61",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c7040ed1-95df-4b06-9744-41b2da53a420","en","ThinkingCap Makes Gemma-4-26B Think 38.9% Less","BottleCap post-trained Gemma-4-26B-A4B-it into ThinkingCap: 38.9% fewer thinking tokens, 0.15pp accuracy change, plus 4.7x throughput on Qwen3.8-27B.","Across twelve benchmarks, accuracy moved by 0.15 percentage points while thinking tokens dropped 38.9%. BottleCap AI calls this post-trained model ThinkingCap. It sits on top of Google's gemma-4-26B-A4B-it, and both the weights and the GGUF quantizations are on Hugging Face, ready to swap in ([original announcement](https:\u002F\u002Fbottlecapai.com\u002Fpost\u002Fthinkingcap-gemma-4-26b\u002F)).\n\nThis is the fifth model in the ThinkingCap series and the first one built on a Google model. The previous four were done for Qwen3.6-27B, Qwen3.8-27B, NVIDIA Nemotron-3.5-Lightning-30B-A3B and Qwen3.8-Flash-Next.\n\n## Whose numbers these are\n\nThese are the vendor's own measurements, and BottleCap is unusually explicit about it in the announcement. The base model was not quoted from published figures: the team ran it themselves through the same evaluation suite, keeping Google's recommended sampling preset unchanged, with thinking enabled, temperature 1.0, top_p 0.95 and top_k 64. Both sides of the comparison, in other words, ran the same race.\n\nThe method is spelled out too. Twelve benchmarks: AIME 2026, two rounds of HMMT, LiveCodeBench v6, GPQA-Diamond, MMLU-Pro, MMMLU, AA-LCR, LongBench v2, RealWorldQA, IFBench and tau2-bench. Random seeds per benchmark range from 4 to 64, with accuracy reported as a mean and a 95% confidence interval. Two failure modes get their own columns: truncation, where a trace hits the generation cap and never closes, and looping, where the model circles the same reasoning forever.\n\nUnder that protocol, eight of the twelve gaps fall inside the run-to-run spread, and five of the twelve actually come out ahead of the base model. Only four move beyond their intervals, and they do not all point the same way: IFBench gains 1.83 points, while GPQA-Diamond loses 0.96, MMMLU loses 0.66 and AA-LCR loses 4.00. That last one is the single real cost in the whole set.\n\n## Where the saved time went\n\nA 38.9% average cut is not spread evenly. Long-context and multimodal sheds the most at 54.4%, knowledge and reasoning 44.4%, maths 36.1%, code 27.6%, while instruction following and agentic give up only 17.2%. The deepest single cut is LongBench v2 at 69.1%; the shallowest is IFBench at 13.7%.\n\nTwo failure rates deserve their own paragraph. Truncation, where the trace hits the 65536-token generation cap and no answer is produced, falls from 4.35% to 0.72%. Looping falls from 3.64% to 0.76%. For anyone running this in production, those two numbers may matter more than the 38.9%.\n\nOne detail deserves caution. The worst accuracy regression, AA-LCR, is also the category where thinking was cut hardest. Long-context reasoning dropped 59.3%, and accuracy dropped 4.00 points to pay for it. The thinking that this model saves is largely taken out of long-context capability. If you are about to ship an agent that leans hard on long-document retrieval, that trade may not be the one you want.\n\n## What actually moves the bill\n\nHalving token spend only gets you half the story. What usually constrains a reasoning model in production is not FLOPs but the KV cache: every concurrent sequence holds a cache that grows with its length, and once memory runs out, the number of tasks that fit in parallel collapses.\n\nBottleCap ran a throughput test in a separate post. Same RTX PRO 6000, 8 GB of KV cache, vLLM 0.29.0, prefix caching disabled, 64 concurrent requests, 20-minute window, one AIME geometry problem. The original Qwen3.8-27B finished 24 tasks; the ThinkingCap version finished 112, a 4.7x difference. Traces were only 2.3x shorter, but average task parallelism rose from 15.1 to 24.9 and token throughput went from 319 to 511 tok\u002Fs. The mechanism is not exotic: shorter traces mean less cache per sequence, so more of them fit on the card, and each finishes sooner, which turns the cache over faster.\n\nTo be precise about it, that 4.7x was measured on a different base model, Qwen3.8-27B, not on Gemma-4. But the Gemma-4 base is arguably the more interesting candidate. Third-party review site CrucibleMark tested its Think mode on a DGX Spark: an overall score of 74.94%, a 91.09-second P95 response time, one timeout out of 49 tasks, and a profile described as a batch DevOps specialist, slow but solid, which is exactly the workload shape that leans on the KV cache. According to that review, the model is a MoE with 25.2B total parameters and roughly 3.8B active per token, released under Apache 2.0, with a 262,144-token context window, and GGUF builds are already available on Hugging Face.\n\n## Who should switch now, and who should wait\n\nIf your service is long-document QA, codebase retrieval or deep research, hold off: those 4 points on AA-LCR will tell you where it hurts first. If the load is batch jobs, background tasks or agent loops, a full order-of-magnitude drop in truncation and looping may be worth more than 0.15 points of accuracy.\n\nTwo practical constraints. First, this is a third-party post-training artifact: the base is Apache 2.0, but the derivative weights carry their own terms, so read the model card on Hugging Face before using it commercially. Second, the PRO version, which the vendor claims shaves up to 87.6% of thinking tokens, is available only through a sales email. There are no public weights.\n\nA third point is the one I keep coming back to. Every ThinkingCap number here comes from the vendor's own evaluation harness, and there is no third-party replication yet. \"Fewer tokens at unchanged accuracy\" has always been the class of claim that most needs independent checking. So run your own A\u002FB on real traffic before treating 38.9% as a fact.\n\nThe real point, then, is not the percentage. Reasoning-model cost optimization is moving from \"make the model think less\" toward \"get more work done per card\", and this company has already laid out both halves of that argument in two separate posts.\n","thinkingcap-gemma-4-26b-shorter-thinking","2026-10-24T11:30:28Z","2026-10-24T11:30:29.274030Z","2026-10-24T11:30:29.274043Z",true,"agent",395,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"8d830f21-a44f-45b8-be6c-eed523d060dc","投机解码的天花板是分词器:3.4M参数换2.4倍","speculative-decoding-bpe-tokenizer-ceiling","2026-11-11T01:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"4f151ed3-74d5-4dc7-99d0-ac03df4c1b08","自一致别整条重采:定位错步,token至多省41%","stepback-error-localized-resampling","2026-11-08T13:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"c643acd4-c460-470b-8e86-4f5bb41479fe","Particle-1:RL 让模型省 45% token","particle-1-rl-token-efficiency","2026-10-29T19:12:56+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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":90,"title":91,"news_slug":92,"published_at":93},"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-m3-1-flash-space-bunny","2026-10-19T21:10:12+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"d6944f35-e3a6-4d5f-a4e1-a546836f7d75","微软基于 Qwen3.5-9B 推出决策模型 Microsoft-Decision-1","microsoft-decision-1-qwen-3-5-9b","2026-10-13T03:00:00+00:00"]