[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-particle-1-rl-token-efficiency":3,"topics-all":38,"news-related-c643acd4-c460-470b-8e86-4f5bb41479fe":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},"c643acd4-c460-470b-8e86-4f5bb41479fe","Particle-1:RL 让模型省 45% token","初创 Particle AI 发布首个模型 Particle-1:在 DeepSeek V4.1 Flash 上做强化学习后训练,官方四个编程与数据工程基准显示平均省约 45% token、任务成本低 34%-52%,pass@1 与基座持平;只调约 1.25% 参数,目前经 API 提供。","推理模型有一个被普遍抱怨的毛病:想得太多,收尾太啰嗦。10 月 28 日,初创公司 Particle AI 发布第一个模型 Particle-1,思路不是做更大的基座,而是用强化学习教同一个基座「把事做完,但少说话」——基座选的是 DeepSeek V4.1 Flash,官方口径是平均少用约 45% 的 token,任务还能更快收尾。\n\n## 省下的都是什么 token\n\n官方在四个基准上对比基座:Terminal-Bench 2.1(终端工作流)、ReactBench(React 应用开发)、ADE Bench(真实 dbt 数据工程)、DeepSWE 1.1(长程软件工程)。输出 token 少 28.4%-52.2%,单位任务成本低 34.4%-51.6%,pass@1 落在基座的置信区间内。合并口径:Particle-1 每任务 59,580 token、0.2211 美元、通过率 72.57%;基座 83,246 token、0.3368 美元、71.15%;对照的 GLM 5.3 Flash 是 0.6523 美元、66.82%。轨迹分析给出了解剖图:三个基准上约九成节省来自推理 token(DeepSWE 上工具调用文本另贡献约三成),最典型的模式是「验证通过就停,不再反复确认」——一个日历任务里,94% 的节省发生在正确文件写完之后。通用能力没塌:GPQA Diamond 90.40% 对基座 90.91%,MMLU 86.08% 对 85.98%。\n\n## 训练配方本身是卖点\n\nParticle 没有重训基座:rank-64 LoRA 只覆盖各层专家,可训练参数约等于全模型的 1.25%;优化器用 GSPO,对整条序列做重要性加权——他们试过的 token 级损失震荡后发散。数据由内部合成系统 Datamaker 生成 Harbor 格式任务,交给 Terminus 2 智能体求解并自动验证;训练栈基于开源 RL 框架 Miles,分两阶段按领域递进。法律合同审查的迁移实验更有意思:适配器强度 1 倍时省 24.6% token,八份报告全部合格;拉到 8 倍省 78.7%,八份只剩三份合格——压缩过头,连有用的限定条件也一起被压掉。\n\n## 所以呢\n\n这件事的信号比这家新公司本身大:模型的竞争轴正在从「分数」扩到「单位任务成本」,而 RL 后训练证明「少想一点」是可以被专门优化的能力。保留意见也要说清:全部数字出自官方自选基准,没有放出权重,目前仅 API 可用,复现要等第三方。对重度使用编程智能体的团队,这种「同分更便宜」的模型值得放进候选池跑一轮。原文见 [Particle 官方博客](https:\u002F\u002Fparticle.ai\u002Fblog\u002Fintroducing-particle-1)。","https:\u002F\u002Fparticle.ai\u002Fblog\u002Fintroducing-particle-1","427e57bf-4bfa-49a9-9e5d-3e4249dff9e9",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",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},"b5fdae94-e9f0-4261-9fff-142a1d152574","en","Particle-1: RL cuts output tokens ~45% on a DeepSeek base","Particle AI's first model RL-post-trains DeepSeek V4.1 Flash to cut tokens ~45%: 28-52% fewer outputs on coding benchmarks, cost down 34-52%, pass@1 flat.","Reasoning models share a widely-mocked habit: they overthink and pad their endings. On October 28, startup Particle AI released its first model, Particle-1 — and instead of building a bigger base, it uses reinforcement learning to teach an existing one to finish the job while saying less. The base is DeepSeek V4.1 Flash, and the official claim is roughly 45% fewer tokens on average, with tasks finishing sooner.\n\n## Where the tokens actually go\n\nParticle's comparison against the base runs on four benchmarks: Terminal-Bench 2.1 (terminal workflows), ReactBench (React application work), ADE Bench (real dbt analytics engineering) and DeepSWE 1.1 (long-horizon software engineering). Output tokens fall 28.4%-52.2%, average cost per task drops 34.4%-51.6%, and pass@1 stays within the base model's confidence intervals. Aggregate view: 59,580 tokens and $0.2211 per task at 72.57% pass rate, versus 83,246 tokens, $0.3368 and 71.15% for the base; GLM 5.3 Flash sits at $0.6523 and 66.82%. Trajectory analysis explains the anatomy: on three benchmarks about 90% of the savings come from reasoning tokens (on DeepSWE, tool-call text contributes around 30%), and the dominant pattern is stop-once-verified — in one calendar task, 94% of the savings came after the correct file had already been written. General capability holds: GPQA Diamond 90.40% vs 90.91% for the base, MMLU 86.08% vs 85.98%.\n\n## The recipe is half the story\n\nParticle didn't retrain the base: a rank-64 LoRA covers the experts in all layers, with trainable parameters equal to about 1.25% of the model. The optimizer is GSPO, applying importance weighting over whole sequences — the token-level loss they tried oscillated and eventually diverged. Training tasks come from Datamaker, an autonomous synthetic-data system, in Harbor format, solved and verified by the Terminus 2 agent; the stack runs on Miles, an open-source RL framework, across two domain-staged phases. A legal-transfer experiment on Harvey LAB contract review adds nuance: at standard 1x adapter strength Particle-1 saves 24.6% of tokens with all eight reports passing; at 8x it saves 78.7%, but only three of eight reports remain complete — over-compression squeezes out useful qualifications along with redundant analysis.\n\n## So what\n\nThe signal here is bigger than one startup: model competition is widening from scores to cost-per-task, and RL post-training shows that thinking less is a trainable capability. Two caveats: the numbers come from the vendor's own benchmark suite, and the model ships API-only with no released weights — third-party replication is pending. For teams running coding agents at volume, a same-accuracy-cheaper model belongs in the candidate pool. Source: [Particle's launch post](https:\u002F\u002Fparticle.ai\u002Fblog\u002Fintroducing-particle-1).","particle-1-rl-token-efficiency","2026-10-29T19:12:56Z","2026-10-29T19:13:01.952515Z","2026-10-29T19:13:01.952526Z",true,"agent",353,[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},"4c9f495a-bf44-4655-8b2b-29b4859a0af3","Ember-1:Kimi K3同质量,少四成token","fireworks-ember-1-kimi-k3-token-efficiency","2026-09-24T23:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"0ecc4f65-2a46-4e7f-8694-510db4f299cf","ThinkingCap:Gemma-4 少想 38.9%","thinkingcap-gemma-4-26b-shorter-thinking","2026-10-24T11:30:28+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a574d32a-c8f7-4d46-ac95-b76406f36b3b","OpenAI 把 GPT-6 拆成两档:Sol\u002FLuna 同款训练法,API 价格砍一半","openai-gpt-6-sol-luna-api-pricing","2026-09-23T00:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"b667e52f-ec7d-4ca4-8d9e-1db81e1a5616","DeepSeek论文:890字节KV缓存的三层架构账","deepseek-v41-flash-kv-cache-paper","2026-09-18T15:10:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"ee62535f-b897-437b-8674-02801632dadb","DeepSeek V4.1 非对称架构首发:读题 8B 答题 16B,KV 缓存砍到初代的 1\u002F437","deepseek-v4-1-flash-ced-kv-cache","2026-09-10T15:10:00+00:00"]