[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-pokee-isaac-28b-10m-context":3,"news-related-b4754043-6b19-499f-8459-f8fc786f4d80":38},{"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},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","Pokee AI 推出 Pokee-Isaac 28B,文本 agent 模型,10M 上下文。RULER 10M 上 93.3%,同基准下 Luna\u002FGemini 3.5 Flash Lite 1M 就 context-overflow。$0.15\u002F$1.00 定价,主打 VPC\u002F本地\u002F端侧部署。","8 月 4 日 Pokee AI 上线 Pokee-Isaac 28B,一个 28B 参数、纯文本的 agent 基础模型,主打一件事:把 agent 跑在客户的边界里,而不是云 API 里。它的上下文窗口拉到 10M tokens——不是只声明「支持」,是在 RULER 长上下文基准每个测量长度下都能跑出可用的分数。\n\n官方技术报告里最刺眼的一张表是 RULER:256K 到 10M,Isaac 全部测出分数,从 96.9 到 93.3。同表里 GPT-5.6 Luna(Azure)到 512K 还撑得住,1M 直接 context-overflow;Gemini 3.5 Flash Lite(Vertex AI)在 1M 拿到 29.4,1M 之后全部归零;Claude Haiku 4.5、Qwen 3.5 122B 在 256K 就已经 0.0;只有 Nemotron 3 Super 自报 91.75 到 1M——是 vendor 自测,不是 Pokee 同环境复测。换句话说,「10M 能用」目前只此一家。\n\n## MRCR v2:多针检索差距拉到 0.3\n\nMRCR v2 是更狠的测试:在长上下文里埋 8 根「针」,要求模型检索并消歧指定的某一根。Isaac 在 256K\u002F512K\u002F1M 三个长度上分别是 0.607\u002F0.743\u002F0.500,Luna 是 0.208\u002F0.173\u002F0.050,Gemini 0.474\u002F0.473\u002F0.205,Haiku\u002FNemotron\u002FQwen 全部 0.000。1M 处 Isaac 对第二名的领先幅度从 256K 的 +0.133 涨到 +0.295,越长的上下文,Isaac 的优势越大。这说明 Isaac 不是「能塞 10M 字符串」,是真的能在 10M 里检索信息。\n\n## Agent 能力:四块基准,两个第一\n\nBFCL v4(AST 评级的 function calling)Isaac 70.94,Luna 70.61,Haiku 67.52——基本是平手,Isaac 0.33 分的领先只是 parity,但意义在「这是能部署在客户边界里那一档」。\n\nτ³-bench(多轮客服任务)Isaac 0.662 四域平均排第一,第二名 Gemini 0.631,Luna 0.527。banking 子项整个面板都跌——全场最好 0.203,Isaac 0.186,几乎算这一档。\n\nTerminal-Bench 2.1(真实 shell)Isaac 65.1%(56\u002F86),Luna 69.8%(60\u002F86)。这是 Luna 唯一赢的一档,差距只有 4 个任务。\n\nMCP-Atlas(36 个 live MCP 服务,不给名字让模型自己发现)Isaac 74.59% 排第三,落后 Luna 3.3 个百分点,但只用了 9.10 turns,而 Luna 用了 14.99。覆盖率差一截,效率高一截。\n\n## 安全:同环境对比,最低 ASR\n\nDTAP red-teaming 测的是 agent 在模拟环境里被注入攻击时的成功率。Isaac 在 Direct ASR、Indirect ASR、Combined ASR 三项全部最低:36.0 \u002F 35.2 \u002F 35.6,Benign Success Rate 82.5。Luna 是 54.4 \u002F 46.1 \u002F 50.1,Gemini 高到 84.1 \u002F 49.5 \u002F 66.3。报告同时说明两条限制:indirect 数据是在 guards-inactive 下测的(因为实验 harness 匹配原生工具名,DTAP 攻击走 MCP);模型只在 1.5% 的恶意任务上明确拒绝——大部分「安全」是「碰巧没动手」而不是「主动挡下来」。这一点 Pokee 自己写得很坦白。\n\n## 服务画像:一卡 B200 撑 137k prefill\n\n单卡 NVIDIA B200 上测出来的:RULER 工作负载下,1M 上下文 TTFT 23.6s、Prefill 42,400 tok\u002Fs、Decode 335 tok\u002Fs;10M 上下文 TTFT 涨到 72.9s,Prefill 反而爬到 137,200 tok\u002Fs,Decode 稳定在 335——prefill 吃长度,decode 不吃。RTX 4090\u002F5090 工作站可跑,Intel Arc Pro B70 上 prefill 是 stock llama.cpp 的 3.6–5 倍,Qualcomm Snapdragon X2 Elite 与 Intel Panther Lake NPU 上也能跑端侧。\n\n定价 $0.15\u002F$1.00(per million in\u002Fout,标记 provisional),给 10M 上下文。Luna 在 1M 之后不再商用定价,Haiku\u002FNemotron\u002FQwen 在 262K 就不能买——「云 API 卖不到的长度」才是 Pokee 的真正锚点。\n\n## 评论:agent 部署范式开始分叉\n\n过去两年的 agent 故事是「更强、更长、更便宜」的云端竞速。Pokee-Isaac 走的是另一条路:用相对克制的 28B 参数换 10M 上下文和 agentic benchmark 平价,把「数据不出边界」做成产品特性,而不是合规补丁。\n\n这件事在监管严苛的行业(医疗、金融、国防、法律、e-discovery)有真实需求——这些场景不是「更喜欢隐私」,是「数据必须留在本地」。当一个 28B 模型能在 RULER 10M 上 93.3%、BFCL v4 平 Luna、agentic benchmark 第一档、ASR 同环境对比最低,意味着客户不再需要为「本地部署」在能力上做大幅妥协——这是一个范式分叉的起点。\n\n来源:[Pokee-Isaac 28B 模型页](https:\u002F\u002Fconsole.pokee.ai\u002Fmodel) | [MarkTechPost 报道](https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F08\u002F08\u002Fpokee-ai-releases-pokee-isaac-28b-a-10m-token-context-agentic-model-built-to-run-inside-the-customer-boundary\u002F)","https:\u002F\u002Fconsole.pokee.ai\u002Fmodel","245abb7f-d909-4acf-971d-931ca5615a42",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"93792da1-471e-449c-861f-66ff3b7a87b0","en","Pokee-Isaac 28B Crams 10M-Token Context Into the Customer Boundary","Pokee AI shipped Pokee-Isaac 28B, a text-only agent foundation model with a 10M-token context window. RULER at 10M scores 93.3%; Luna and Gemini 3.5 Flash Lite context-overflow at 1M. Priced $0.15\u002F$1.00 per million in\u002Fout tokens, licensed for VPC, on-prem, or on-device deployment.","On August 4, 2026, Pokee AI shipped Pokee-Isaac 28B — a 28B-parameter, text-only agent foundation model built around a single thesis: run agents inside the customer's boundary, not against a cloud API. Its context window stretches to 10M tokens — and, more importantly, it produces usable scores at every measured length on the RULER long-context benchmark.\n\nThe most striking table in the technical report is RULER from 256K to 10M. Isaac returns 96.9, 96.7, 95.0, 95.8, 96.7, 93.3 — a usable score at every length. GPT-5.6 Luna on Azure holds up to 512K and context-overflows at 1M. Gemini 3.5 Flash Lite on Vertex AI scores 29.4 at 1M and zeros out beyond. Claude Haiku 4.5, Qwen 3.5 122B, and Nemotron 3 Super all hit 0.0 by 256K, except Nemotron's vendor-self-reported 91.75 at 1M — which was measured by NVIDIA, not Pokee, on the same installation. In other words, \"usable at 10M\" currently has exactly one implementation.\n\n## MRCR v2: the multi-needle gap widens to 0.3\n\nMRCR v2 is the harder test: eight needles buried in long context, the model must retrieve and disambiguate a specified one. Isaac scores 0.607 \u002F 0.743 \u002F 0.500 at 256K \u002F 512K \u002F 1M. Luna is 0.208 \u002F 0.173 \u002F 0.050. Gemini is 0.474 \u002F 0.473 \u002F 0.205. Haiku, Nemotron, and Qwen all sit at 0.000. The Isaac lead over the second-best baseline grows from +0.133 at 256K to +0.295 at 1M — the longer the context, the bigger Isaac's edge. This isn't \"fits 10M tokens\" — it's \"retrieves reliably inside 10M tokens.\"\n\n## Agent capability: four benchmarks, two first places\n\nOn BFCL v4 (function calling graded by AST and state-transition matching) Isaac scores 70.94, Luna 70.61, Haiku 67.52 — effectively a tie; the 0.33-point Isaac lead is parity, not a decisive margin. But the practical significance is: \"you can have this on your own boundary at all.\"\n\nτ³-bench (multi-turn customer service against an evolving simulated user): Isaac averages 0.662 across four domains, ranking first. Gemini is 0.631, Luna 0.527. On the banking sub-task the entire panel collapses — the best score in that column is 0.203, Isaac 0.186 — which is essentially a wash.\n\nTerminal-Bench 2.1 (real shell, real test suite): Isaac 65.1% (56\u002F86), Luna 69.8% (60\u002F86). This is the one benchmark where a cloud baseline wins — and the gap is four tasks.\n\nMCP-Atlas (discover and compose tools across 36 live MCP servers, neither server nor tool named): Isaac 74.59% in third place, behind Luna by 3.3 points, but using 9.10 turns per task versus Luna's 14.99. Lower coverage, materially better trajectory length.\n\n## Security: lowest ASR in same-installation comparison\n\nDTAP red-teaming places the agent in simulated environments and measures whether injected attacks succeed. Isaac records the lowest scores on every axis: Direct ASR 36.0, Indirect ASR 35.2, Combined ASR 35.6, with 82.5 Benign Success Rate. Luna is 54.4 \u002F 46.1 \u002F 50.1. Gemini is 84.1 \u002F 49.5 \u002F 66.3. Two limits the report states plainly: indirect is \"guards-inactive\" because the harness matches native tool names while DTAP attacks arrive over MCP; and explicit refusals fired on only 1.5% of malicious tasks — most of the current safety is incidental rather than declined. Pokee writes this honestly.\n\n## Serving profile: 137k prefill on one B200\n\nOn a single NVIDIA B200 under the RULER workload: at 1M context, TTFT 23.6s, Prefill 42,400 tok\u002Fs, Decode 335 tok\u002Fs; at 10M context, TTFT grows to 72.9s, Prefill climbs to 137,200 tok\u002Fs, Decode holds flat near 335 — prefill scales with length, decode doesn't. Single-GPU workstation serves on RTX 4090 or RTX 5090. On Intel Arc Pro B70, prefill is 3.6–5× stock llama.cpp on the same card. On-device on Qualcomm Snapdragon X2 Elite and Intel Panther Lake NPU.\n\nPricing is $0.15 per million input tokens \u002F $1.00 per million output, against 10M context. Luna is not sold commercially above 1M. Haiku, Nemotron, and Qwen all cap at 262K. \"Lengths you can't buy from a cloud API\" is the actual anchor.\n\n## Commentary: an agent deployment fork is opening\n\nThe last two years of agent stories have been a cloud race — stronger, longer, cheaper, all hosted. Pokee-Isaac takes the other branch: trade a moderate 28B parameter count for 10M context and agentic benchmark parity, and turn \"data doesn't leave the boundary\" from a compliance patch into a product feature.\n\nThe demand is real in regulated industries — healthcare, financial services, defense, legal, e-discovery. These are not customers who \"prefer privacy\" — they are customers whose data is not permitted to cross an external API boundary at all. When a 28B model can hit 93.3% on RULER at 10M, tie Luna on BFCL v4, rank first on agentic benchmarks, and record the lowest combined ASR on DTAP, then \"on-prem\" no longer requires a big capability concession. That is a fork.\n\nSources: [Pokee-Isaac 28B model page](https:\u002F\u002Fconsole.pokee.ai\u002Fmodel) | [MarkTechPost coverage](https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F08\u002F08\u002Fpokee-ai-releases-pokee-isaac-28b-a-10m-token-context-agentic-model-built-to-run-inside-the-customer-boundary\u002F)","pokee-isaac-28b-10m-context","2026-08-20T14:00:00Z","2026-08-20T07:17:15.663941Z","2026-08-20T07:17:15.663950Z",true,"agent",73,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"5878a668-282c-4b88-b2b8-7eef40b7938c","LFM2.5-2.6B：2.5GB 内存跑本机 Agent 220 tok\u002Fs","lfm2-5-2-6b-on-device-agent","2026-08-11T00:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"79c1684f-f61d-4799-b3d0-6450c4ad10e8","Muse Glimmer:Meta 把 30B 「常驻本地的智能体」开源,把 Agent 拉到笔记本里 7×24 跑","muse-glimmer-30b-open-agentic-local","2026-08-10T00:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"5ef8c5da-7221-4e22-bd46-c1ef55c3120d","韩国 Solar Open 2：Linear Attention 进 MoE，250B\u002F15B 激活","upstage-solar-open-2","2026-07-23T03:30:00+00:00"]