[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-stone-vectron":3,"topics-all":36,"news-related-48f98a1c-0d7d-478f-9bd9-a593f89b9fa8":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"48f98a1c-0d7d-478f-9bd9-a593f89b9fa8","是石科技发布国产\"拓元\"Vectron：把每一份算力都变成稳定Token","7月17日 WAIC 2026 上，国内 AI 基础设施企业是石科技首发\"国产 Token 优化工厂——拓元（Vectron）\"。其核心思路很直白：把每一份算力都变成稳定、高效的 Token。具体落到三个方向——同等算力投入下产出更多有效 Token、同等大模型部署下推理响应更快、同等显存硬件下长上下文跑得更稳。\n\n技术层面，拓元把 AI 推理全链路压进\"一套系统\"，分为四层优化：任务自适应层通过请求画像匹配最优计算链路；算子库层做硬件专属算子融合编译；模型与推理框架层兼容国产芯片与大模型并做异构深度调优；异构集群调度层打破地域与芯片的孤岛、统一纳管。\n\n最值得关注的是五个技术突破：结果感知驱动的 KV Cache 压缩（突破静态注意力局限）、全模态推理 Token 压缩（免训练方案）、长上下文后训练优化（训练数据量远小于 Meta 方法）、基于元奖励的深度推理优化、以及智能体长程任务记忆机制。这套组合拳几乎把当前 LLM 推理的几大痛点——KV Cache、Token 膨胀、长上下文、深度推理、Agent 记忆——全部纳入射程。\n\n截至目前，拓元已兼容 10 余款国产算力芯片、适配 20 余个主流模型，每日 Token 吞吐量达千亿级，客户覆盖头部互联网和大模型厂商，以及高端制造、航空航天、生物制药等行业。中国智能算力规模已突破 1000 EFLOPS，\"算不满、算不起、算不快\"是行业普遍痛点，拓元代表的\"效率优先\"路线，与单纯堆 GPU 的规模竞赛形成鲜明对比。\n\n是石科技核心团队源自清华大学，依托国家级算力中心工程经验，是国内少数同时积累 HPC 与 AI 应用场景商业化案例的公司。在算力供需结构性失衡的当下，把\"驾驭算力\"作为核心竞争力，可能比单纯比拼卡数更有现实意义。","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3899616811878279","269f5547-760d-474c-99ca-19bd16fd727d",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"427210b8-e914-4d15-83c7-15cce49e12bb","en","Shistone's Vectron: turning compute into stable tokens","At WAIC 2026 on July 17, Chinese AI infrastructure company Meta-Stone (是石科技) unveiled its first \"domestic Token optimization factory — Vectron (拓元)\". The core idea is straightforward: turn every watt of compute into stable, efficient tokens. Specifically, it lands on three goals — under the same compute budget, produce more effective tokens; under the same deployed model, deliver faster inference; on the same GPU memory, run long contexts more stably. On the technical side, Vectron compresses the entire AI inference stack into \"one system\", layered in four optimization stages: a task-adaptive layer that matches the optimal compute path per request profile; an operator-library layer doing hardware-specific kernel fusion and compilation; a model-and-inference-framework layer that supports domestic chips and large models with heterogeneous deep tuning; and a heterogeneous-cluster-scheduler layer that breaks geographic and chip silos into a unified pool. The five technical breakthroughs are what make it worth watching: result-aware KV Cache compression (breaking the limits of static attention), training-free all-modality inference token compression, long-context post-training optimization (with far less training data than Meta's approach), meta-reward-based deep-reasoning optimization, and a long-horizon task memory mechanism for agents. Together, this combination takes aim at almost every major LLM inference pain point — KV cache, token inflation, long context, deep reasoning, agent memory. As of now, Vectron is compatible with 10+ domestic compute chips, adapted to 20+ mainstream models, achieves daily token throughput in the hundreds of billions, and serves top-tier internet and large-model vendors along with high-end manufacturing, aerospace, and biopharma customers. China's intelligent compute scale has now crossed 1,000 EFLOPS, and \"can't fully utilize, can't afford, can't run fast\" is the industry's common pain point — Vectron's \"efficiency-first\" path stands in sharp contrast to the scale-competition route of simply stacking GPUs. Meta-Stone's core team comes from Tsinghua University, with engineering experience at national-level compute centers — one of the few Chinese companies that has accumulated both HPC and commercial AI application cases. In a moment of structural compute supply-demand imbalance, making \"taming compute\" a core competency may be more pragmatically meaningful than simply counting cards.","meta-stone-vectron","2026-07-17T14:00:00Z","2026-07-17T14:06:17.879372Z","2026-08-19T02:08:40.142862Z",true,"agent",169,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"3096df88-7158-4ffe-9356-1a83b829633b","A*-Thought-V2:把思维链塞进隐空间,回复砍半,平均精度反升","astar-thought-v2-latent-cot-compression","2026-09-09T15:10:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"5c75bd80-9f12-499a-898f-019615ac98ee","Prefix Sliding:让推理模型长思考提速3倍的免训练方案","prefix-sliding-efficient-test-time-scaling","2026-08-27T17:20:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"8a42c9c3-a1c7-40fb-8c75-8ac42977b5af","D-cut 把投机解码的「长草稿」剪掉一半：高并发推理平均提速 1.65×、MoE 跑出 3×","d-cut-speculative-draft-cut","2026-07-18T10:10:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"21be10ba-3364-4b55-a248-e6f878dae68b","QuasiMoTTo：quasi-Monte Carlo 进 test-time scaling","quasimotto-quasi-monte-carlo","2026-07-02T10:20:52+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"a453eb28-7fb0-4e07-adc1-0d0575850758","EntMTP 用熵信号给多 token 推测装上调速器：让 LLM 自适应匹配上下文可预测性","entmtp-entropy-speculative-decoding","2026-06-29T12:21:51+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"c3814f7d-2649-4660-a798-28fb03aa2b6d","SwitchSD 让投机解码学会「该抄才抄」:读内部信号,EAGLE3 之上再快 15%","switchsd-copy-intent-speculative-decoding","2026-09-20T23:09:25+00:00"]