New Meta paper shows, small models may not be bad predictors of scale; they may just be getting under-tuned. Finds scaling laws emerge around 4M parameters, where models can train in under 1 hour on 1 GPU. Small models are unusually sensitive to hyperparameters. With 4 or 16 configurations per scale, the law is basically invisible; at 64 it appears but extrapolates poorly, and at 256 it becomes accurate. As models get larger, good settings occupy more of the search space, while the effective number of hyperparameters near the optimum drops toward 1. That helps explain why scaling laws look cleaner at larger sizes: the models are easier to tune. As a check, small-scale runs recover that pre-norm transformers scale better than post-norm over the tested range. There is a limit: extrapolate too far beyond the measured scales, and statistical errors can dominate. For model research, cheap experiments may need more tuning breadth, not more model size. - arxiv. org/abs/2608.11859 Title: "Small-Scale Experiments: Are They There Yet?"

译Meta 新论文发现，小模型可能只是欠调优，而非规模扩展的差预测器。规模法则在约 4M 参数处涌现，该规模模型可在 1 块 GPU 上 1 小时内完成训练。每规模需 256 组超参数配置时法则才准确，小模型对超参数异常敏感。

Rohan Paul@rohanpaul_ai · 14小时前31

If you want diverse model behavior, give people control below the prompt layer.

译杨立昆重申，AI技术必须广泛开放共享，如同印刷术和互联网一样赋能人类。他认为社会需要价值观、语言和专长各异的多样化AI系统，而这只能通过开放基础模型实现。他赞同扎克伯格对"AI危险论导致权力集中"的批评，并称与Anthropic CEO关于"多AI制衡"的争论中，开放路线是唯一出路。

Rohan Paul@rohanpaul_ai · 1天前36

A new Meta FAIR paper finds a blind spot in the Chinchilla scaling law that becomes expensive when you extrapolate. Chinchilla can look almost perfect inside a training grid and still mispredict what happens at the frontier. The problem is: Chinchilla assumes model size and training data help separately, but the experiments show that each changes how useful the other one is. Skaling adds just 1 extra term to capture that connection, cutting prediction error by about 1.5-3× and getting full-grid Chinchilla-level prediction accuracy with roughly 10× less profiling compute. On Farseer, the difference becomes huge at frontier scale: at 2×10^25 FLOPs, Chinchilla points to ~380 tokens per parameter, while Skaling and the paper's direct estimates land around 20-40. - arxiv. org/abs/2608.07222 Title: "Skaling: Chinchilla's Exponents Meet Kaplan's Coupling"

译Meta FAIR 新论文发现 Chinchilla 缩放定律在外推至前沿规模时存在盲点，因其假设模型大小与训练数据独立作用。新方法 Skaling 仅增加一项耦合项，将预测误差降低约 1.5-3 倍，并以约 10 倍更少的 profiling 算力达到全网格精度。

Rohan Paul@rohanpaul_ai · 1天前55

Bloomberg: Alibaba says Qwen crossed 3B global downloads across its expanding open-weight model family. Further for downstream reuse number, Hugging Face counted 151,448 Qwen-based derivatives, 2.6x Meta's total footprint and 4.7x Llama repositories. "Qwen has become part of the default workflow for developers deciding what models to fine-tune and deploy"

译彭博社：阿里巴巴表示，其不断扩大的开源权重模型家族 Qwen 全球下载量已突破 30 亿次。此外，在下游复用数量方面，Hugging Face 统计到 151，448 个基于 Qwen 的衍生模型，是 Meta 总规模的 2.6 倍，是 Llama 仓库数量的 4.7 倍。 "Qwen 已成为开发者决定微调和部署何种模型的默认工作流程的一部分。"

Rohan Paul@rohanpaul_ai · 2天前47

Meta's new paper, the dangerous failure mode is not just a wrong judge, but a correct judge that can be persuaded into becoming wrong. We are increasingly using AI models to judge other AI models. But what if the AI being judged can simply argue with the judge until the judge changes its decision? Meta tested exactly that. Across 9 frontier models, an adversarial LLM could flip judge verdicts on 62-91% of tested cases under sustained adaptive persuasion. And changing the judge's mind usually didn't fix a mistake. It made the judgment worse: under the adaptive attack, 70% of successful flips moved away from the ground truth. That creates a very practical problem for agent systems. If one AI is supervising another AI, the supervised agent may eventually be able to contest, negotiate with, or strategically persuade its own evaluator. - arxiv. org/abs/2608.12645

译Meta 新论文揭示 AI 评审系统的危险失败模式：被评审的 AI 可通过持续自适应说服，在 9 个前沿模型上翻转 62-91% 的评审判决。更严重的是，改判通常不是修正错误--70% 的成功翻转反而偏离了真实答案，对智能体监督体系构成实际威胁。

elvis@omarsar0 · 2天前46

Brilliant new paper from Meta. LLM judges get validated on accuracy against golden data. That says nothing about whether the verdict survives when questioned. The Wiggle Framework stress-tests 9 frontier models across 14 judging tasks along three axes, stability under re-prompting, stability under a single challenge, and stability under sustained pressure. They find that every model wiggles. Verdicts flip 25 to 71% of the time under static pushback, and 62 to 91% against an adversarial persuader. Pressure that changes a judge's verdict is almost always net-corrupting against ground truth. Baseline jury majority strength turns out to be the best single-shot predictor of which items will move. Paper: https://arxiv.org/abs/2608.12645 Track more trending AI papers in our academy: https://academy.dair.ai/

译Meta 发布 Wiggle 框架，对 9 个前沿模型在 14 项裁判任务中施压测试，发现所有模型都会"摇摆"：静态反驳下改判率 25-71%，对抗性说服下高达 62-91%。改变裁判判决的压力几乎总是损害其相对真实答案的准确性，而基线陪审团多数优势是预测哪些项目会变动的最佳单一指标。

Rohan Paul@rohanpaul_ai · 2天前36

The Meta/Oxford study finds, a multimodal model may need surprisingly little image-generation data if language and visual understanding are trained with it from the start. So, you probably don't need to spend that much training compute teaching a multimodal model to generate images. Language training already helps the model with vision, and learning to understand images also makes it better at generating them. But training it to generate images does very little for language or image understanding. That leads to a very uneven training mix. In their 1T-token experiments, the best overall split was 70% language, 25% image understanding, and only 5% image generation. They then tested this at 13.5B scale over 2T tokens. Even with 5x fewer image-generation tokens than the balanced setup, GenEval improved from 0.467 to 0.482, while language and image understanding improved too. There's another lesson: don't bolt vision on too late. The longer the model trains only on language, the more it starts ignoring the image and relying on language shortcuts. The authors call this "vision laziness." - arxiv. org/abs/2608.05000 Title: "Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes"

译Meta 与牛津大学研究发现，若语言与视觉理解同步训练，多模态模型仅需极少量图像生成数据。1T token 实验中最佳配比为 70% 语言、25% 图像理解、5% 图像生成；13.5B 规模 2T token 测试中，图像生成 token 减少 5 倍，GenEval 仍从 0.467 提升至 0.482。研究还指出过晚引入视觉会导致"视觉惰性"，模型倾向依赖语言捷径。

Chubby♨️@kimmonismus · 4天前10

I was live on @etnshow , it was a great pleasure. I still get incredibly nervous in front of the camera, but I hope I was able to say some interesting things about Grok and Meta's release.

译我刚刚在 @etnshow 上直播，非常荣幸。面对镜头我依然会非常紧张，但我希望我关于 Grok 和 Meta 发布的一些观点还算有趣。

OpenRouter@OpenRouter · 4天前精选72

The first open-weights model from @AIatMeta Superintelligence Labs, Muse Glimmer is live on OpenRouter! It's a 30B dense text+image model with an Apache 2.0 license, built to be a reliable local agent w/ scores of 75.5 MCP Atlas and 51.2 SWE-Bench Pro. https://openrouter.ai/meta/muse-glimmer-30b

译Meta AI 超级智能实验室的首个开放权重模型 Muse Glimmer 已在 OpenRouter 上线！这是一款 30B 密集文本+图像模型，采用 Apache 2.0 许可证，旨在成为可靠的本地智能体，MCP Atlas 得分 75.5，SWE-Bench Pro 得分 51.2。 https://openrouter.ai/meta/muse-glimmer-30b

X.PIN@thexpin · 4天前50

Manus says it will return to operating as an independent company. As part of the transition, some user data created on or after Dec. 29, 2025 - the day Meta announced its acquisition of Manus - will be deleted on Aug. 23-24. Future data will be stored in the U.S. and Singapore. Manus says new features will again push the limits of general AI agents. But without the backing of a major foundation-model provider, whether it can deliver on that ambition remains an open question.

译Manus 表示将恢复为独立公司运营。作为过渡的一部分，2025 年 12 月 29 日（Meta 宣布收购 Manus 之日）及之后创建的部分用户数据将于 8 月 23 日至 24 日被删除。未来数据将存储在美国和新加坡。Manus 称新功能将再次突破通用 AI 智能体的极限。但在没有大型基础模型提供商支持的情况下，能否实现这一雄心仍是未知数。

Rohan Paul@rohanpaul_ai · 5天前32

Today's edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/meta-released-muse-glimmer-a-30b 🗞️ Meta Superintelligence Labs released Muse Glimmer, a 30B-parameter model built for always-on local agents on one consumer GPU. 🗞️ Mark Zuckerberg just dropped a really long piece on Meta's vision for the AI future where superintelligence is available to everyone. 🗞️ Bernie Sanders demands OpenAI, Anthropic, and Meta pause AI development over escalating safety concerns. 🗞️ Qwen released a multimodal tool layer for AI agents. 🗞️ Australia's first known autonomous AI agent driven hack. 🗞️ Metis: Memory Foundation Model

译Meta Superintelligence Labs 发布 Muse Glimmer，一款 30B 参数模型，专为单张消费级 GPU 上的常驻本地智能体设计。同日，扎克伯格发文阐述 Meta 让超级智能惠及所有人的 AI 愿景，伯尼·桑德斯则呼吁 OpenAI、Anthropic 和 Meta 暂停 AI 开发。

Rohan Paul@rohanpaul_ai · 5天前42

Meta's open-sourced Muse Glimmer 30B vs Muse Spark 1.2 Glimmer delivered 5 runnable games at roughly one-fifth Muse Spark's cost. Very Interesting experiment by @aimlapi . the setup: • Muse Glimmer 30B - via aimlapi[.]com. cost: $0.02 • Muse Spark 1.2 - via aimlapi[.]com. cost: $0.11 Glimmer still made mistakes, especially in the racing game, and Spark produced better-looking graphics. Meta built Glimmer as a 29.6B dense model distilled from Muse Spark for local agentic and coding work. Its model card lists 131K+ context and quantized versions under 20GB, sized for 24GB or 32GB consumer hardware.

译Meta 开源 Muse Glimmer 30B，在 aimlapi 上以 $0.02 成本一次生成 5 个可运行游戏，约为 Muse Spark 1.2（$0.11）的五分之一。Glimmer 在 Flappy Bird 上表现更优，但赛车游戏仍有失误且画质较粗糙。该模型为 29.6B 稠密模型，由 Muse Spark 蒸馏而来，支持 131K+ 上下文，量化版小于 20GB。

Rohan Paul@rohanpaul_ai · 5天前40

Meta's new open-sourced Muse Glimmer-30B beats Gemma 4 31B across 3 arcade-game prompts. Test was done on atomic[.]chat (@atomic_chat_hq), a desktop app that runs LLMs locally. Each model ran locally on a separate RTX 5090 with 32GB VRAM Muse Glimmer's games play like the originals. Space Invaders sped up as aliens died, Tetris locked pieces and cleared lines, and Arkanoid's mechanics worked without crashing. So Muse paid heavily for that result: 83.4K tokens and 17.7 minutes, versus Gemma's 17.8K tokens and 4.5 minutes.

译Meta 开源 Muse Glimmer-30B 在三个街机游戏生成任务中击败 Gemma 4 31B。测试在 Atomic Chat 桌面应用本地运行，每模型各用一块 RTX 5090（32GB VRAM）。Glimmer 游戏还原度高，但代价高昂：耗 83.4K tokens、17.7 分钟，而 Gemma 仅 17.8K tokens、4.5 分钟。

Aidan Gomez@aidangomez · 6天前29

This is the right vision for the future of AI, focused on progress and building a better future instead of obsession over risks, doom, and downside. There's a better future to be brought about by those who believe in it.

译这是 AI 未来的正确愿景，聚焦于进步和建设更美好的未来，而非执着于风险、厄运和负面。相信美好未来的人，终将缔造它。

Chubby♨️@kimmonismus · 6天前42

What a day: - Meta is back in the open-source game, delivering Glimmer, an excellent agentic mid-sized model, and will soon release Spark 1.2 with open weights. - Anthropic tried to solve the Riemann Hypothesis with an unreleased Claude model. It failed, but along the way unexpectedly improved a longstanding lower bound for the proportion of zeros of the Riemann zeta function known to satisfy the hypothesis, from 41.6% to 67.2%. - At the same time, Sonnet's prices were cut for the long term, surely because of the competition, and the news around Gemini 3.5 Pro seems to suggest that Google is moving straight to Gemini 4. All in one day. Crazy. I'm going to sleep now. See you tomorrow!

译Meta 重回开源，发布智能体模型 Glimmer，并将开源 Spark 1.2。Anthropic 用未发布 Claude 模型尝试证明黎曼猜想未果，但意外将黎曼 zeta 函数零点满足猜想的比例下界从 41.6% 提升至 67.2%。同时 Sonnet 长期降价，Gemini 3.5 Pro 消息暗示 Google 或直接推进至 Gemini 4。

Artificial Analysis@ArtificialAnlys · 6天前同事件78

Meta returns to open weights: Muse Glimmer, its first open-weights release since Llama 4, scores 35 on the Artificial Analysis Intelligence Index. It is a 30B-parameter model, and the first from Meta to be released under Apache 2.0 Muse Glimmer (high) arrives 16 months after Llama 4, scoring 21 points above Llama 4 Maverick (14), Meta's last open weights release. It sits alongside Kimi K2.5 (Reasoning, 36) and just behind Qwen3.6 27B (Reasoning, 38) and Ling 3.0 Flash (38), and creates a two-tier Meta lineup together with the proprietary flagship Muse Spark 1.2 (xhigh, 57) @AIatMeta shared access with us ahead of public release for benchmarking. Congratulations to @AIatMeta, @finkd, and @alexandr_wang on the release! Key Takeaways: ➤ Meta's open-weights line is back, under its most permissive license yet. Every prior Meta open release shipped under a Llama License; Muse Glimmer uses Apache 2.0, placing almost no restrictions on commercial use or derivatives ➤ Strong intelligence for its parameter count. At 30B parameters, Muse Glimmer scores 5 points above Gemma 4 31B (Reasoning, 30) at the same size, and effectively matches 1T total parameter Kimi K2.5 (Reasoning, 36) with 33x fewer parameters. Qwen3.6 27B (Reasoning, 38) remains ahead on the Intelligence vs Parameters frontier at a slightly smaller size ➤ Small enough to self-host on a single GPU, even at full context. Muse Glimmer is a 30B dense model (including a ~1.8B vision encoder) with weights at ~60 GB in BF16 and ~18 GB in 4-bit. It features a hybrid-attention mechanism with three sliding-window layers for every global layer, which holds KV cache memory use to ~1.8 GB (minimum) at its pre-extension 128K context. This means the model can run at full context on a single H100 at BF16 precision, or on a higher-spec MacBook or RTX 5090 at 4-bit, with more breathing room if the vision encoder is not required ➤ Agentic knowledge work is its weakness relative to its size class. Muse Glimmer scores 953 Elo on GDPval-AA v2, below the 1,000 human baseline and behind other models at its intelligence level, including the similarly sized Qwen3.6 27B (Reasoning, 1141), and Gemini 3.5 Flash-Lite (1141). Knowledge calibration follows the same pattern: its AA-Omniscience Index of -33 is low for its intelligence level, driven by an 82% hallucination rate (Qwen3.6 27B: 49%) rather than accuracy, where it matches its peers. Agentic tool use is the exception, with Muse Glimmer scoring 24% on Tau3-Banking, ahead of Gemini 3.5 Flash-Lite (18%) and Qwen3.6 27B (17%), among the best in its class Other model details: ➤ Parameters: 30B Dense ➤ Context window: 128K tokens, plus extension ➤ License: Apache 2.0 ➤ Openness Index: 44, indicating relatively high openness/transparency vs. other major open-weights models ➤ Pricing and output speed: at the time of release, Meta is not serving the model on their API; pricing and serving speed are TBC depending on third-party providers

译Meta 发布开源模型 Muse Glimmer，这是其自 Llama 4 以来首个开源权重模型，30B 参数，在 Artificial Analysis 智能指数上得分 35，也是 Meta 首个采用 Apache 2.0 许可的模型。

Andrew Ng@AndrewYNg · 6天前47

Thank you Mark, Alex and the whole Meta team for your contributions to open weight AI.

译感谢 Mark、Alex 以及整个 Meta 团队对开放权重 AI 的贡献。

Yuchen Jin@Yuchenj_UW · 6天前46

That Llama is back! Exciting to see Meta is back in open-source AI. @finkd: "Open source is a positive and important force for empowering people and preventing centralization that is detrimental for both safety and the economy." Hope more frontier labs like Anthropic and OpenAI embrace open source/open weights as the path to a positive AI future.

译Meta 宣布开源 Muse Glimmer，一个 30B 参数的稠密模型，可本地运行；并即将发布最新基础模型 Muse Spark 1.2 的权重。扎克伯格称开源是赋能用户、防止权力集中的积极力量。作者对此表示欢迎，并希望 Anthropic、OpenAI 等前沿实验室也能拥抱开源。

elvis@omarsar0 · 6天前36

Impressive new paper from Meta. (bookmark it) Scaling laws assume model size and training data act on loss independently. This work introduces Skaling law, which couples capacity and data through a single interaction exponent. The extra term cuts mean absolute percentage error by 1.5x to 3x across both interpolation and extrapolation. The largest corrections land in the data-scarce and heavy-overtraining regimes where the standard Chinchilla and Kaplan forms drift. Paired with a sparse grid restricted to low-compute runs, it extrapolates the full grid using roughly 10x less compute than a uniform sweep. Why does it matter? Deployment now happens well past compute optimal. A law that stays accurate there, and that can be fit from small runs, changes how a pretraining budget gets planned. Paper: https://arxiv.org/abs/2608.07222 Track more trending AI papers in our academy: https://academy.dair.ai/

译Meta 新论文提出 Skaling law，通过单一交互指数耦合模型容量与训练数据，将平均绝对百分比误差在插值与外推场景下降低 1.5x 至 3x。该定律在数据稀缺和过度训练区间修正了标准 Chinchilla 与 Kaplan 形式的偏差，并可用约 10 倍更少算力从稀疏网格外推完整训练配置。

AYi@AYi_AInotes · 6天前同事件57

Meta 刚扔了个王炸,开源了一个 30B 的 agent 模型,Apache 2.0,单张消费级 GPU 就能跑,工具调用、多步推理、出错自己恢复,不用上云不用 API key,7×24 在你电脑上跑,小扎这波真的很牛逼, 这个模型叫 Muse Glimmer, 30B 参数,dense 模型,Apache 2.0,权重直接放, 本地能跑。而且很快,Muse Spark 1.2-- 他们最新的基础模型 -- 权重也要放出来。 Alexandr Wang 带的 Superintelligence Labs 做的。 @ylecun LeCun 在下面回了句:好棋。确实是好棋啊, 但不是开源万岁那种好棋, 先跟大家说说 Glimmer 到底是什么。这不是又一个聊天模型, 它专门为本地 agent 工作流优化的:工具调用,多步推理,失败恢复,多模态理解。量化之后,单张消费级 GPU,24G 或者 32G 显存,就能跑。也就是说,一个能自己调工具、自己从错误里恢复、7 乘 24 在你电脑上跑的 agent,不需要上云,不需要 API key,一张显卡搞定,真的有点屌炸, Meta 之前的 Muse Spark 系列是走封闭路线的, 那是 Meta 第一次正经做 proprietary。现在突然把 30B 的 agent 模型放出来, 还预告基础模型也要开。从商业的角度,我觉得这不是小扎回心转意了, 老哥肯定是算过账的, 当你的封闭模型还没建立起足够壁垒的时候, 开源是最好的武器。用一个中等规模但 agent 能力强的模型, Apache 2.0,无负担商用,快速占住开发者的本地部署生态。 @Alibaba_Qwen 在抢,Gemma 在抢,@DeepSeek 在抢。 Meta 如果继续闭源,开源话语权就没了。现在他们想用 Glimmer 抢本地 agent"这个心智, 用 Spark 1.2 抢基础模型生态。其实就是战略的一环,不是啥情怀, 更值得注意的是Glimmer 主打的不是对话质量,也不是 benchmark 分数,是 agentic reliability。规划,工具调用,出错了能自己恢复。这说明 Meta 已经把能真正完成任务作为下一阶段的核心指标了。不是你问它答得多好,是给它一个目标它能不能自己干完。我觉得这个转向比参数规模重要得多, 对做 agent 的开发者来说,这是一件大事。以前本地跑的 agent 模型要么太小不够用, 要么太大跑不动。 30B dense,专门为 agent 优化,单卡能跑,Apache 2.0。这就是本地 agent 时代的一张入场券啊, 以后不用什么都往云上扔了, 所以别只看 Meta 又开源了, 看他们开源的是什么。一个为你电脑上 7 乘 24 运行的 agent 设计的模型, 从最新基础模型蒸馏出来的专用模型, 用最宽松许可放出来抢生态的武器。开源在这里不是信仰, 是棋盘上的一步, 而且下得很准。

译Meta 开源 30B 参数的 Muse Glimmer，Apache 2.0 许可，专为本地常驻 Agent 工作流优化，支持工具调用、多步推理与失败恢复，量化后单张 24G/32G 消费级 GPU 即可运行。

elvis@omarsar0 · 6天前42

Given how much Meta has done for open-source AI, it only felt right that they get back in the game. This is a huge statement from Meta and a big deal for the AI ecosystem. More than that, it points to a future where it's important to own your intelligence.

译Meta 宣布开源 30B 参数稠密模型 Muse Glimmer 权重，可本地运行，并预告将发布最新基础模型 Muse Spark 1.2 的权重。Meta 此举被视为对开源 AI 生态的重大表态，也预示着"拥有自身智能"的未来趋势。

elvis@omarsar0 · 6天前39

"Open source is a positive and important force for empowering people" Mark understands this more than anyone. Really happy that Meta is committing again to strengthening the open source ecosystem.

译"开源是赋能人们的积极而重要的力量" Mark 比任何人都更懂这一点。很高兴看到 Meta 再次承诺加强开源生态系统。

Rohan Paul@rohanpaul_ai · 6天前47

"In the coming weeks, we are also going to open the weights for Muse Spark 1.2, our latest foundation model and one of the leading models in the world." • Mark Zuckerberg Also a personal superintelligence available 24/7 to everyone.

译扎克伯格发文阐述Meta AI愿景，宣布将在未来几周开源最新基础模型Muse Spark 1.2的权重，并称其为全球领先模型之一。他还提出面向所有人的24/7个人超级智能，并呼吁前沿AI实验室向政府共享中间训练检查点。

Nathan Lambert@natolambert · 6天前22

Llama 5 let's go thanks @alexandr_wang

译Llama 5 走起，感谢 @alexandr_wang

Ethan Mollick@emollick · 6天前47

Spark is the big news and is a good model. Not quite at the frontier of open models from China, and still well behind the closed frontier, but the best non-Chinese open weights model released in a year. Of course, a lot depends on continuing to release new open models to keep up

译Meta 开源 Muse Glimmer（30B 参数稠密模型，可本地运行），并将发布 Muse Spark 1.2 权重。Ethan Mollick 评价 Spark 是近一年最佳非中国开源权重模型，但仍未达中国开源模型前沿，且远落后于闭源前沿。

Rohan Paul@rohanpaul_ai · 6天前同事件58

Mark Zuckerberg just dropped a really long piece on Meta's vision for the AI future where superintelligence is available to everyone. some strong quotes from the blog. • "Any policy that slows American model releases -- even by a month -- could add significant risk to American leadership while letting foreign models race ahead." • "For example, if a self-improving AI system focused on optimizing its compute efficiency, it could theoretically invent ways to squeeze 100x or more intelligence out of each gigawatt." • "There is no such thing as a singular benevolent superintelligence." • "While there are risks to releasing capable models, the most dangerous scenario from this perspective would be leading AI labs training powerful models and keeping them for themselves." • "I also propose that frontier AI labs should share intermediate training checkpoints of new models for government use and review rather than waiting until training has completed."

译扎克伯格发布长文阐述 Meta 的 AI 愿景，主张超级智能应惠及所有人。他警告任何延缓美国模型发布的政策（哪怕仅一个月）都可能危及美国领导地位，并提议前沿 AI 实验室向政府共享中间训练检查点以供审查。

🚨 AI News | TestingCatalog@testingcatalog · 6天前同事件52

BREAKING 🔥: Meta released Muse Glimmer, a new open-weight 30B model under the Apache 2.0 license! Muse Spark 1.2 is planned to be released as open-weight soon as well. Outperforms Gemma 4 31B and Qwen 3.6 27B on loads of benchmarks. Muse Glimmer can be quantized to under 20GB with room for the KV cache, perception encoder, and a drafter. Meta is back for real 👀

译Meta 发布开源权重模型 Muse Glimmer，为 30B 参数规模，采用 Apache 2.0 许可证，可在 24GB 显存上运行且不损失智能体可靠性。该模型在多项基准测试中超越 Gemma 4 31B 和 Qwen 3.6 27B，量化后体积可压缩至 20GB 以下。Meta 同时宣布将很快开源 Muse Spark 1.2。

Rohan Paul@rohanpaul_ai · 6天前同事件52

Meta Superintelligence Labs released Muse Glimmer, a 30B-parameter model built for always-on local agents on one consumer GPU. weights on HuggingFace under an Apache 2.0 license, • At full precision the model would need 55GB+ of memory. • Meta compressed the weights to roughly 4-bit precision, shrinking the language model to under 20GB. That leaves room inside a 24GB or 32GB envelope for the KV cache, a perception encoder for screenshots, and a speculative decoding drafter. • The compressed model still leads Gemma4-31B and Qwen3.6-27B on MCP Atlas, at 75.5 against 54.2 and 62.5. That lead reverses on computer use and terminal coding, where Qwen3.6-27B scores higher on OSWorld-Verified, TerminalBench 2.1, and SWE-Bench Verified. • Instead of writing one word at a time, Muse Glimmer lets a small helper model called DFlash guess a whole block ahead. The main model then checks that block in one pass, keeping the correct guesses and fixing the wrong ones. • On an RTX-5090, that lifted decode speed from 74.9 to 233 tokens per second, a 3.1x gain measured through llama.cpp. • Apple hardware gains less, 1.8x on an M5-Max and 1.5x on an M4-Max, reaching 50 and 38 tokens per second.

译Meta Superintelligence Labs 发布 Muse Glimmer，一款 30B 参数、专为单消费级 GPU 上的常驻本地智能体设计的模型，权重以 Apache 2.0 协议开源。

Chubby♨️@kimmonismus · 6天前57

I'm genuinely happy that Meta is back in the open-source game. The weights for Spark 1.2 will be released soon as well! I just wish they had kept the name Llama.

译Meta 回归开源，其 30B 参数模型 Muse Glimmer 在 24 项基准中 12 项取得最佳成绩，并在 19 项上超越 Gemma4-31B、14 项超越 Qwen3.6-27B。该模型主打智能体能力，4-bit 量化版仅约 17GB，性能损失约 1%。权重将以 Apache 2.0 协议完全开放，Spark 1.2 权重也将很快发布。

clem 🤗@ClementDelangue · 6天前18

Meta is back! well done @finkd @alexandr_wang!

译Meta 回来了！干得漂亮 @finkd @alexandr_wang！

Chubby♨️@kimmonismus · 6天前62

Huge: Meta says it will resume releasing open-source AI models "soon" as part of a much larger plan: delivering personal superintelligence to billions of people! Zuckerberg commits to free or affordable access, personal agents with a private mode even Meta cannot inspect, and alignment centered on each user's goals rather than Meta's values. He also proposes independent board approval of model-release safety criteria, and early government access to intermediate training checkpoints. "Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about. Your agent will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more." Meta commits to free or affordable access, with additional compute allocated through a dynamic auction. It also promises a fully private mode in which even Meta cannot see or grant access to personal information. Superintelligence is within reach guys. The golden ear of scientific discovery infront of us and Open Source the way to get there!

译Meta 宣布将"很快"恢复发布开源 AI 模型，作为向数十亿人提供个人超级智能计划的一部分。扎克伯格承诺免费或可负担的访问、Meta 也无法查看的完全隐私模式，以及以用户目标而非 Meta 价值观为中心的对齐。他还提议由独立董事会批准模型发布安全标准，并让政府提前访问中间训练检查点。

Chubby♨️@kimmonismus · 6天前同事件52

LFG: Meta's Muse Glimmer is a remarkably capable model for its size: 30B parameters, local deployment, and the best reported result on 12 of 24 benchmark rows against Gemma4-31B and Qwen3.6-27B. By my count, it beats Gemma on 19 of 24 rows and Qwen on 14. Its strongest area is agentic work: -MCP Atlas: 75.5 vs. Qwen's 62.5 -DeepSearch QA: 74.6 vs. 71.1 -τ3-Banking: 23.5 vs. 16.7 -SWE-Bench Pro: 51.2 vs. 50.2 -AA-LCR: 80.0 vs. 73.3 Qwen remains ahead on OSWorld, TerminalBench and most multimodal tests. Gemma leads the two primary safety metrics. The 4-bit model fits into roughly 17GB with only 1% average degradation across 15 benchmarks, according to Meta. And this is an unusually open Meta release: public, ungated weights under Apache 2.0, including BF16 and quantized versions, the perception encoder and DFlash drafter. Meta delivered! Really excited for this one!

译Meta 发布开源模型 Muse Glimmer（30B 参数，支持本地部署），在 24 项基准中 12 项取得最佳成绩，对比 Gemma4-31B 胜 19 项、对比 Qwen3.6-27B 胜 14 项，最强领域为智能体任务（MCP Atlas 75.5 vs 62.5）。

Alexandr Wang@alexandr_wang · 6天前25

personal superintelligence should be available to everyone, and opening access to our models is abig part of that. read more from mark: http://meta.com/futureisforeveryone

译个人超级智能应该惠及每一个人，开放我们模型的访问权限正是其中至关重要的一环。更多内容请阅读马克的文章：http://meta.com/futureisforeveryone

Alexandr Wang@alexandr_wang · 6天前同事件71

excited to be releasing open weights for muse glimmer today, a 30b model that runs on a single consumer gpu, with open weights for a version of muse spark 1.2 coming soon. two very different models, both headed into people's hands, with more to come.

译Meta 今日开源 Muse Glimmer，一个可在单张消费级 GPU 上本地运行的 30B 参数稠密模型，并即将开源最新基础模型 Muse Spark 1.2 的权重。Meta 表示坚定支持开源，Alexandr Wang 与 MSL 团队参与了模型开发。

Alexandr Wang@alexandr_wang · 6天前精选77

1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon. we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability. 🧵

译1/ 今天有个重大消息：我们很快将发布 Muse Spark 1.2 的开源权重版本。同时，我们还将发布 Muse Glimmer--一个 30B 参数的智能体模型，采用 Apache 2.0 协议开源权重。Muse Glimmer 可在 24GB 显存上运行，且不损失智能体可靠性。🧵

Mark Zuckerberg@finkd · 6天前精选69

I believe everyone should have access to superintelligence, and I wrote a long piece about Meta's philosophy and values for building a positive future for everyone. http://meta.com/thefutureisforeveryone

译我相信每个人都应能使用超级智能，我撰写了一篇长文，阐述 Meta 为所有人构建积极未来的理念与价值观。http://meta.com/thefutureisforeveryone

Mark Zuckerberg@finkd · 6天前同事件60

Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

译今天，我们还开放了 Muse Glimmer 的权重，这是一个出色的 30B 参数稠密模型，可以在本地运行。很快，我们还将发布最新基础模型 Muse Spark 1.2 的权重。Meta 是开源的坚定支持者，我为这些发布感到自豪。祝贺 @alexandr_wang 和 MSL 团队在这些模型上的出色工作。

AI at Meta@AIatMeta · 6天前精选58

Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on consumer hardware like a Mac or PCs with performant GPUs. In keeping with our long tradition of sharing fundamental AI research, we're releasing model weights under a permissive Apache 2.0 license. 🧵👇

译推出 Muse Glimmer，一款开放权重、300 亿参数的模型，专为本地、常驻运行的智能体工作流优化。与同尺寸领先模型相比，Muse Glimmer 在关键智能体用例和基准测试中表现出色，并设计为完全在消费级硬件（如 Mac 或配备高性能 GPU 的 PC）上运行。秉承我们长期分享基础 AI 研究的传统，我们以宽松的 Apache 2.0 许可证发布模型权重。

Rohan Paul@rohanpaul_ai · 6天前54

Meta's new retrieval paper is a reminder that better language models do not necessarily require more language generation. It asks a useful question: why generate item IDs at all when the LLM can learn the retrieval space directly? Instead of autoregressively generating item IDs or ranked lists, the paper puts a 0.6B Qwen3 model inside a classic two-tower retriever, so item embeddings stay precomputable and retrieval remains a fast vector search. The stronger cross-encoder is used as a teacher, transferring its candidate-ranking distribution into the cheaper two-tower student. That distillation is doing a lot of the work: removing it cuts Recall@10 by 13.3%, 23.1%, and 8.0% on Beauty, Sports, and Toys. Against OneRec-Think, which uses Qwen3-8B, the 0.6B two-tower model reports higher Recall@10 on all three public datasets, although its NDCG results are mixed. On Meta's internal production data, the same approach matches a heavily tuned DLRM retriever's normalized entropy while using just 0.5% of the training data. Those production results are self-reported, but the engineering implication is specific. For first-stage retrieval, the LLM does not necessarily need to generate anything; semantic representations inside a two-tower architecture may be a much more practical use of the model. - arxiv. org/abs/2607.25346 Title: "The Case Against Generation for Retrieval: Discriminative Language Models as Effective Retrievers"

译Meta 新论文提出用判别式语言模型做检索，将 0.6B Qwen3 模型嵌入经典双塔架构，替代自回归生成 item ID，使向量检索保持高速。更强的交叉编码器作为教师进行知识蒸馏，移除蒸馏后 Recall@10 在 Beauty、Sports、Toys 上分别下降 13.3%、23.1%、8.0%。

elvis@omarsar0 · 7天前37

New research from Meta. Agent harnesses are still mostly authored by hand. This makes it hard to tune robust agent harnesses for long-horizon tasks. In this new work, agents learn harness policies offline and deploy them to construct and update external harness state online during runtime task execution. EvoHarness-RL learns that policy instead. Belief, Progress, and Experience are exposed as harness state the policy can act on. Supervised harness fine-tuning teaches the action space, then cost-aware GRPO explores when to read, update, and consolidate during a long run. Qwen3-8B reaches 96.9% on ALFWorld. Two dynamics come out of the training. Harness annealing means recurring harness-use patterns get absorbed into the model policy, and the agent shifts from frequent calls toward selective access. Harness evolution means progress updates and experience consolidation compress the workspace into a compact task-adaptive state. This shows that long-horizon agents get more from a trainable coordination policy than from bigger tools or larger memories. Paper: https://arxiv.org/abs/2608.05446 Track more trending AI papers in our academy: https://academy.dair.ai/

译Meta 新研究提出 EvoHarness-RL，让智能体离线学习工具编排（harness）策略，并在运行时在线更新外部状态，替代手工编写。Qwen3-8B 在 ALFWorld 上达到 96.9% 准确率。训练中出现的"工具编排退火"与"工具编排进化"两种动态表明，长程任务智能体从可训练的协调策略中获益，胜过更大的工具集或记忆。

8月17日

8月16日

8月15日

8月14日

8月12日

8月11日

8月10日

全部 AI 动态