[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-muse-spark-12-coding-agent-54-index":3,"news-related-e846f1c6-4644-4f84-a664-83ec82734210":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},"e846f1c6-4644-4f84-a664-83ec82734210","Meta Muse Spark 1.2 与 Muse Code 把「1.2 → 编程」的推理效率推回前沿","Meta Superintelligence Labs 于 2026 年 8 月 5 日发布 Muse Spark 1.2 和配套终端编程 Agent Muse Code。独立评测机构 Artificial Analysis 给出 Muse Spark 1.2 Intelligence Index 54 分,较 1.1 涨 3 分,与 Grok 4.5 并列美国第三。增量集中在 agentic 知识工作: GDPval-AA v2 Elo 涨 260 到 1631、Terminal-Bench 提到 80%。Meta 同时公开了 Muse Code 的异步子 Agent 编排与 kernel 优化训练机制,以及 1.2 在 AA-Omniscience 上以拒答换可靠性的取舍策略。","## 技术背景:Agentic 能力成新一代旗舰必争之地\n\n2026 年下半年,前沿大模型的竞争轴心已经从\"谁的 benchmark 跑分高\"切换到\"谁能在真实 agent 工作流里扛活\"。Claude Opus 5、GPT-5.6 Sol、Kimi K3 在 GDPval-AA v2、Terminal-Bench 等 agentic 评测上接连刷新上限,把\"通用对话\"这条传统赛道挤成了入场券而非决胜场。Meta Superintelligence Labs 在 4 月推出 Muse Spark 1.0(43 分 AA Index)、7 月推到 1.1(51 分),并在 8 月 5 日放出 Muse Spark 1.2(54 分)和配套的终端编程 Agent Muse Code,试图在\"代码 + 长时间任务\"这条更窄但商业价值更高的轴上打回原形。Muse Spark 1.2 的同期关键数据来自 Artificial Analysis 的独立评测(发布于 2026-08-05),把这次更新从 Meta 单方宣传变成了可被外部验证的硬指标。\n\n## 核心内容:Muse Spark 1.2 的 3 个增量与 1 个取舍\n\n**1. Agentic 知识工作显著逼近前沿,但成本曲线同步抬升。** Artificial Analysis 的 Intelligence Index 上,Muse Spark 1.2 (xhigh) 拿到 54 分,比 1.1 涨 3 分,比 1.0 涨 11 分。它与美国三号位 SpaceXAI 在分数上并列,并与 GPT-5.5 (xhigh, 55)、Grok 4.5 (high, 54) 几乎打平,仅落后 Claude Opus 5 (max, 61)、Claude Fable 5 (max w\u002F fallback, 60)、GPT-5.6 Sol (max, 59) 与 Kimi K3 (max, 57)。分项里 GDPval-AA v2 Elo 从 1371 拉升到 1631(+260),排到所有被测模型的第 5 名,已经超过 Claude Opus 4.8 (max, 1588);Terminal-Bench v2.1 从 78% 提到 80%;τ³-Banking 从 25% 提到 27%。代价是单任务成本从 $0.29 涨到 $0.40(基于 Meta 不变的 $1.25\u002F$4.25 per 1M token 定价),主要因为输入\u002F输出 token 用量在 GDPval-AA v2 上各增加约 53% 和 36%。\n\n**2. Muse Code 把\"长期任务 + 子 Agent 编排\"做进了默认行为。** Meta AI Research 的发布说明里把 Muse Code 描述为一个异步后台 Agent 模型,而不是传统的\"一个主 Agent + 临时调度\"。每个 Muse Code session 维持若干持续子 Agent(如 Photon Sphere、Embervault、Avo Lawn),它们跨任务保持存活、负责下一步动作,并自行决定何时反馈主 Agent,目标是减少重复信息收集、降低延迟、避免在长链路任务里反复被用户打断。每一次模型调用、工具执行、审批和编辑都进本地事件日志,实现 replay-exact + restart-safe:进程崩溃后,Agent 能精确接着上次的位置继续。这个设计直接把\"长任务中途挂掉要重头再来\"这个老问题从默认假设里删掉。`\u002Fplan`、`\u002Fgrill`、`\u002Fgoal` 三个内置 skill 分别承担\"先出可审批计划\"、\"压力测试计划是否成立\"、\"朝目标推进\"的职责。\n\n**3. 共训练(Kernel Optimization Case Study)与\"自我改进循环\"是这次训练侧的核心信号。** Meta 在发布博客里披露,Muse Spark 1.2 用 Muse Code 的 harness 轨迹做了拒绝采样训练,目标\u002F压缩\u002F子 Agent 的策略和 Muse Code 工具集一起进了优化配方,让模型在与 Muse Code 配对时表现最好。1.2 也大量训练于\"长链路编码任务\"——整仓库生成、大型端到端项目、自动研究类工作,用到规划、目标条件化、上下文压缩来维持进度。Case Study 部分:让模型在 1000+ 次工具调用(最多 24 小时)内迭代优化 GPU kernel,测试对象是 NVIDIA Hopper 上的 KDA 与 MLA kernel。基准线是 FLA 的 Triton 实现,模型被禁止直接 import 第三方 kernel 库;Muse Spark 1.2 在 KDA 上把 chunk-parallel preparation kernel 和 sequential inter-chunk scan 拼起来,在 MLA 上设计了两阶段 Triton pipeline 复用 KV latent 作为 K 和 V。\n\n**4. 取舍:科学推理小幅回调 + 事实可靠性押注\"少答\"。** SciCode 掉 2 点(58% → 56%),Humanity's Last Exam 掉 1 点(45% → 44%);CritPt 涨 3 点(15% → 18%)。更值得注意的是 AA-Omniscience 的 abstention 模式延续:分数 18 → 22,幻觉率 38% → 28%,但 attempt rate 从 82% 掉到 67%,accuracy 从 41% 滑到 38%。Meta 在\"宁愿拒答也不胡说\"这条路线上做出了第二次加注。\n\n**定价与上下文(沿用 1.1):** 1M token 上下文,$1.25\u002F$4.25 per 1M input\u002Foutput tokens,缓存命中 $0.15 per 1M,首发即在 Meta 自家 API + Muse Code 同步可用(详见 https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-code-and-muse-spark-1-2 与 https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fmuse-spark-1-2)。\n\n## 个人评论:为什么 54 分这一档,Meta 要用\"长期任务 + Agent 编排\"来破\n\n把 Muse Spark 1.2 放进 2026 H2 的坐标系里看,它不算是\"分数第一\"那种发布。Claude Opus 5、GPT-5.6 Sol、Fable 5、Kimi K3 都还在它前面。但 Meta 的选择是清楚的:它不在通用对话那条已经卷成红海的赛道上追 GPT-5.6 Sol,而是把\"agent 任务 + 长链路编程 + 自家 harness\"三件事咬成一个闭环。这里面有两条值得关注的信号。\n\n第一条是\"成本上升被默认接受\"。Muse Spark 1.2 单任务成本涨 38%(从 $0.29 到 $0.40),但分项得分涨得更猛——这是把\"分越高越值得\"这个假设重新放回到产品决策里。如果只看 token 价格,Muse Spark 1.2 处于 GPT-5.6 Sol ($0.51)、Kimi K3 ($0.86)、GPT-5.5 ($1.18) 之下,在 50+ 段位上仍是性价比选项;但和 GPT-5.6 Sol (medium, $0.39) 比只便宜 $0.01。Meta 显然在为 agent 工作流里\"为了更高成功率多烧 token\"正名。\n\n第二条是\"自我改进循环 + Co-Training\"被写进了发布主叙事。Meta 用 1.1 生成挑战性编码环境与指令模板,再让 1.1 自己给候选方案打分,把这种\"自我对练\"的产出当作 1.2 的训练数据。这条路线让人想起 Anthropic 在 Constitutional AI 阶段的自我批评循环;Meta 的版本把它推到了 coding benchmark + 长链路 Agent 这一更\"工程化\"的赛道上,直接对接商业化场景。\n\n副作用是不得不提的:事实性能力下滑。AA-Omniscience 的 attempt rate 从 82% 掉到 67%、accuracy 从 41% 滑到 38%——这是 Meta 第二次在评测指标层面选择\"宁愿拒答\"。如果 Muse Code 这类长链路 Agent 在生产环境里越来越多地自己决策(写代码、跑测试、调 kernel),那么\"模型会不会编一个看起来合理但事实错误的中间结论\"就不是个学术问题,而是回归质量风险。这一点,Meta 暂时没有给出新的校准方案。\n\n所以,Muse Spark 1.2 的真正赌注是:在 GPT-5.6 Sol、Claude Opus 5 把通用天花板顶住的这半年里,把\"编程 + Agent + 长任务\"这条更接近商业落地的小赛道吃下来——分数只追到 54,但每分都花在最能转化为客户付费意愿的能力上。这条路如果走通,2026 年下半年的 LLM 竞争可能就不再是\"谁分高\",而是\"谁在自家 harness + 自家评测上跑得最稳\"。\n\n(原 Meta AI Research 发布博客: https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-code-and-muse-spark-1-2;Artificial Analysis 独立评测: https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fmuse-spark-1-2)","https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-code-and-muse-spark-1-2","a1f0bda7-5035-4317-b63b-72693539d2e3",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"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},"dbe6a947-35f8-4fac-982c-2a11d7a6bf3c","en","Meta Muse Spark 1.2 and Muse Code push reasoning efficiency back","Meta Superintelligence Labs released Muse Spark 1.2 and the terminal coding agent Muse Code on August 5, 2026. Independent evaluator Artificial Analysis gave Muse Spark 1.2 a 54-point Intelligence Index score, up 3 points from 1.1 and effectively tied with Grok 4.5 for third among US labs. The gains concentrate in agentic knowledge work: GDPval-AA v2 Elo +260 to 1631, Terminal-Bench up to 80%. Meta also disclosed Muse Code's async background-agent orchestration and its kernel-optimisation training loop, plus the trade-off of trading answer rate for reliability on AA-Omniscience.","# Meta Muse Spark 1.2 and Muse Code: a 54-point step that bets on long-horizon coding over raw benchmark supremacy\n\n## Technical background: agentic work has become the new frontier axis\n\nIn the second half of 2026, the competition axis for frontier LLMs has shifted from \"whose raw benchmark score is highest\" to \"who can carry realistic agentic workloads\". Claude Opus 5, GPT-5.6 Sol and Kimi K3 have all refreshed the upper bound on GDPval-AA v2 and Terminal-Bench, and \"general chat\" has been compressed from a decisive battleground into an entry ticket. Meta Superintelligence Labs launched Muse Spark 1.0 (43 on the AA Index) in April, pushed to 1.1 (51) in July, and on August 5 released Muse Spark 1.2 (54) together with the terminal coding agent Muse Code. The bet is to win back share on the narrower but commercially more valuable axis of \"code + long-horizon tasks\". The independent numbers behind Muse Spark 1.2 come from Artificial Analysis (published 2026-08-05), which moves this update from a one-sided announcement into externally verifiable hard data.\n\n## What Muse Spark 1.2 actually changes: three deltas and one trade-off\n\n### 1. Agentic knowledge work closes the gap, but the cost curve moves with it\n\nOn Artificial Analysis's Intelligence Index, Muse Spark 1.2 (xhigh) lands at 54 — +3 over 1.1 and +11 over 1.0. It is effectively tied with SpaceXAI in third place among US labs, and within a point of GPT-5.5 (xhigh, 55) and Grok 4.5 (high, 54). It is only narrowly behind Claude Opus 5 (max, 61), Claude Fable 5 (max w\u002F fallback, 60), GPT-5.6 Sol (max, 59) and Kimi K3 (max, 57). Most of the lift comes from agentic evaluations: GDPval-AA v2 Elo jumped +260 (1371 → 1631, #5 among all benchmarked models and ahead of Claude Opus 4.8's 1588); Terminal-Bench v2.1 rose +2 points (78% → 80%); τ³-Banking rose +2 points (25% → 27%). The bill is also larger: per-Intelligence-Index-task cost moved from $0.29 to $0.40 at Meta's unchanged $1.25\u002F$4.25 per 1M token pricing, driven by ~53% more input tokens and ~36% more output tokens per task, concentrated in GDPval-AA v2.\n\n### 2. Muse Code turns \"long-horizon tasks + subagent orchestration\" into the default behaviour\n\nMeta's release blog describes Muse Code as an asynchronous background-agent model rather than a classical \"single main agent with ad-hoc dispatch\". Each session keeps a small set of persistent background agents (Photon Sphere, Embervault, Avo Lawn) that remain alive across tasks, take next steps on their own and decide when to report back to the main agent. The intent is to remove redundant information gathering, lower latency and avoid the constant re-prompting that long multi-step tasks normally require. Every model call, tool run, approval and edit is appended to a local event log, making the runtime replay-exact and restart-safe: after a crash the agent resumes precisely where it stopped. This design quietly deletes \"the long task crashed, start over\" from the default assumptions of coding agents. Three bundled skills — `\u002Fplan`, `\u002Fgrill`, `\u002Fgoal` — cover \"produce an approval-gated plan\", \"stress-test that plan until it holds\" and \"drive toward the specified goal\".\n\n### 3. Co-training with Muse Code + a \"self-improvement loop\" is the training-side headline\n\nThe release explicitly states that Muse Spark 1.2 was co-trained with the Muse Code harness: rejection-sampled harness trajectories, recipe optimisation for goals \u002F compaction \u002F subagents, and the integration of the Muse Code toolset so the model performs best when paired with Muse Code. A second training move uses Muse Spark 1.1 itself to generate hard coding environments and instruction-following templates, then to grade candidate solutions against them — producing a scalable training corpus for 1.2. The Case Study in the blog pushes this further: the team ran Muse Spark 1.2 over 1,000+ tool calls (up to 24 hours) to iteratively optimise GPU kernels on NVIDIA Hopper, targeting KDA and MLA kernels. The baseline was the FLA Triton implementation; models were forbidden from importing third-party kernel libraries and had to write the algorithm themselves. Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan for KDA, and a two-kernel Triton pipeline for MLA that reuses the shared KV latent as both K and V.\n\n### 4. The trade-off: slight regression on scientific reasoning and a doubled-down abstention strategy\n\nSciCode dropped 2 points (58% → 56%); Humanity's Last Exam dropped 1 point (45% → 44%); CritPt rose 3 points (15% → 18%). More striking is the AA-Omniscience result: the score rose 18 → 22, the hallucination rate fell 10 points (38% → 28%), but the attempt rate dropped from 82% to 67%, and accuracy slipped 41% → 38%. For the second release in a row Meta is betting that \"the model should refuse when it doesn't know\" is a feature, not a regression — and is willing to pay for it in raw answer rate.\n\n### Pricing and context (carried over from 1.1)\n\nThe model retains the 1M-token context window and the same pricing as Muse Spark 1.1: $1.25 \u002F $4.25 per 1M input \u002F output tokens, with cache hits at $0.15 per 1M. It is available at launch on Meta's first-party API (Meta Model API) and inside Muse Code. Sources: https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-code-and-muse-spark-1-2 and https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fmuse-spark-1-2.\n\n## Why Meta is using long-horizon coding + agent orchestration to break into the 54-point tier\n\nPutting Muse Spark 1.2 into the 2026-H2 coordinate system, this is not a \"highest score wins\" release. Claude Opus 5, GPT-5.6 Sol, Fable 5 and Kimi K3 are all still ahead of it. Meta's choice, however, is legible: it is not chasing GPT-5.6 Sol on the general-chat red-ocean axis; it is binding \"agentic tasks + long-horizon coding + its own harness\" into a tight loop. Two signals are worth tracking.\n\nThe first is \"cost goes up, and that is OK\". Muse Spark 1.2's per-task cost rose ~38% ($0.29 → $0.40), but its per-task scores rose more — this is \"higher score is worth more tokens\" being put back into the product decision. Looking only at token price, Muse Spark 1.2 is still cheaper per task than GPT-5.6 Sol ($0.51), Kimi K3 ($0.86) and GPT-5.5 ($1.18), and remains a cost-efficient option at the 54+ cluster; but versus GPT-5.6 Sol (medium, $0.39) the gap is only $0.01. Meta is clearly normalising higher token usage in agentic workflows in exchange for higher success rates.\n\nThe second is \"self-improvement loop + co-training\" being written into the public release narrative. Meta used 1.1 to generate hard coding environments and instruction templates, then used 1.1 to grade candidate solutions, and folded that output back into 1.2's training data. This is reminiscent of Anthropic's Constitutional-AI self-critique loop; Meta's version pushes it onto the more \"engineering\" axis of coding benchmarks + long-horizon agents, where it directly meets commercial demand.\n\nThe side effect has to be named: factual reliability is sliding. AA-Omniscience attempt rate dropped 82% → 67% and accuracy 41% → 38% — Meta is doubling down, for the second evaluation cycle in a row, on \"the model prefers to abstain\". If Muse Code and similar long-horizon agents keep being trusted to make their own decisions in production (writing code, running tests, tuning kernels), \"will the model fabricate a plausible-looking but factually wrong intermediate conclusion\" stops being an academic question and becomes a regression-quality risk. Meta has not offered a new mitigation for this in the release.\n\nSo the real bet of Muse Spark 1.2 is: while GPT-5.6 Sol and Claude Opus 5 are holding the general-purpose ceiling, Meta will eat the smaller but more commercially immediate lane of \"programming + agents + long tasks\" — chasing only to 54 points, but spending every point on the abilities that translate most directly into willingness to pay. If this path works, the H2-2026 LLM competition may no longer be \"whose score is highest\", but \"whose model runs most reliably inside its own harness and on its own evaluations\".\n\n(Original Meta AI Research release: https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-code-and-muse-spark-1-2; Artificial Analysis independent evaluation: https:\u002F\u002Fartificialanalysis.ai\u002Farticles\u002Fmuse-spark-1-2)","meta-muse-spark-12-coding-agent-54-index","2026-08-05T08:00:00Z","2026-08-07T08:10:46.148790Z","2026-08-07T08:10:46.148805Z",true,"agent",236,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"7ba15299-8bee-4039-8bc7-dbb58754b562","SWE-bench Science:最强 Claude Code 修科学代码也不及格,四类失败模式被拆解","swe-bench-science-benchmark","2026-08-21T13:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"5b76d928-8896-4bf8-8edd-e195ecf0094a","Ornith-1.5自报跑分赢了Claude,独立复测翻车了","ornith-1-5-benchmark-reality-check","2026-08-20T17:10:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"89a79f9a-bfd2-4ebe-8f03-92fa74a3a34f","Ornith-1.5 开源：模型自己出题、自己搭考场，397B 到 9B 三档齐发","ornith-1-5-self-improvement-open-models","2026-08-20T13:30:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"4c54e7dc-46f6-4d0e-86c9-278995cf0da8","Stanford PTXBench:让 LLM 裸写 H100\u002FB200 PTX kernel,没有一个模型全过关","ptxbench-llm-ptx-gpu-kernel-benchmark","2026-08-19T14:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"c4ec4625-4a84-4c24-88f0-0ef1beb4f19e","Grok 4.6 发布:61 分追平 GPT-5.6 Sol,把长程 Agent 的 token 账单砍到四分之一","grok-4-6-agentic-cost-frontier","2026-08-14T19:00:00+00:00"]