Serving Drama 3 on open vLLM infrastructure means teams can self-host expressive TTS instead of depending on hosted APIs, and the 128-stream per-H200 figure gives a concrete capacity baseline for cost planning. Developers building voice agents or audio pipelines can run it today with the one-line serve command on upstream vLLM. Fish Audio competes directly with hosted TTS providers on throughput while keeping voice cloning local.
NVIDIA 预告下周 GTC 上可能发 Nemotron 3.5 Super VL，一个带视频理解的多模态模型，架构细节已经泄露出来了。
Skywork 把世界模型做成了真能玩的游戏，单张 5090 跑 60 FPS，改代码就能换玩法，权重代码全开源了。
A sparse mixture-of-experts design that keeps only 2B parameters active per token lets a 24B model run locally while matching a larger dense-activation rival on tool-use work, changing what is practical on consumer hardware. Developers building local agent workflows can pull it with a single Ollama command or grab the open weights on Hugging Face to fine-tune. Compared with Qwen3.6-35B-A3B, it targets the same agentic niche at lower active-parameter cost.
The update closes specific gaps analysts flagged weeks earlier, which is a rare case of a vendor visibly acting on public feedback. Developers building image and video pipelines can now get object coordinates directly from the model instead of adding a separate detection step. The 262K context also makes long multilingual document and media workflows viable on a single model.
有人实测 DeepSeek 服务恢复后的新模型，速度 180 t/s，表现接近 Claude，可能是 4.1 Pro，想去试的可以看看。
Ant 旗下 Ling 3.1 Flash 放出开源权重了，本地跑就能用，同一提示词下生成效果赢过 GPT-6 Luna，还便宜 2.8 倍，想省钱本地部署的可以试试。
有人把 DeepSeek Flash 拆开算了笔账：单任务 0.13 美元，还搬出 OpenRouter 的 token 用量数据佐证它已经是全球用得最多的模型，价格敏感的开发者可以看看。
有人改了《哈克贝利·费恩历险记》原文喂给 9 个模型测理解力，只有 Claude Fable 5 两次都过关，Gary Marcus 拿它说事，看看你用的模型表现如何。
阶跃星辰把 Step-5-Preview 开源了，600B 参数只用激活 27B，输入输出都能吃 100 万上下文，长文档场景可以试试。
Serving efficiency, not raw scale, is where V4.1 Flash wins: it delivers 1.3x to 6.2x the throughput per MW of V4 Pro 0813 across Blackwell hardware while scoring higher on intelligence. Teams running agentic coding workloads can use the AA-AgentPerf data to pick hardware — B300 nodes stay within about 20% of rack-scale GB300 NVL72, so single 8-GPU nodes are viable for this model. As an open-weights release from DeepSeek, it widens cheap self-hosting options for Southeast Asian inference providers competing with rented frontier APIs.
物理评测的标准答案自己都错了，GPT-6 Astra 反而是对的，最后拿 98%，这条挺有意思的。
Meta 新模型 Muse Wonder (Max) 的实测排名出来了，智能体任务比上一代强不少，但价格贵了一倍多，文本榜冲到第 2，数据都在这。
The result measures long-tail factual recall plus 3D generation quality in a single offline benchmark, a combination few evals test. Builders comparing frontier models on spatial reasoning and world knowledge can use the same sandboxed prompt setup for their own side-by-side runs. Claude Opus 5.5 (Anthropic) edged out Grok 4.7 (xAI) on geographic accuracy of the name placement.
有人吐槽物理基准 CritPt 被 Astra 刚发布就刷爆，顺便夸了 OpenAI 的 RL 训练管线，聊前沿模型饱和速度的可以看看。
一个才 18M 的 OCR 模型，CPU 就能跑，每秒吞 49 页，做截图取字功能正合适，表格和手写就别指望它了。
The MI355X's 288GB of HBM3E per GPU gives it an edge on very large models like the 1.6T-parameter DeepSeek V4 Pro, but serving-stack maturity on ROCm remains the bottleneck at higher loads. Teams evaluating accelerator purchases for agentic inference can now compare AMD and NVIDIA on realistic coding-agent workloads rather than synthetic throughput numbers. All results run on ROCm nightly builds of vLLM and SGLang, so scores should improve as the software stack matures.
Meta entering the agentic model space puts it in direct competition with OpenAI's operator-style agents and Anthropic's computer-use models, rather than just chat assistants. Developers building personal-assistant or workplace automation products should read the researcher blog to compare benchmark scores against their current agent stack before choosing a base model.
OpenRouter 上架了 Jev 1.14，输出免费输入才 $0.042/M，还加了看图能力和多语言改进，性价比党可以试试。
Kombaico 做了个测 AI 设计创造力的榜单，GPT-6 Astra 才拿 65.5 分，人类是 92.3，模型们设计得都挺像，论文和代码都开源了。
Google 的 Embedding Gemma 2 冲上 Hugging Face 热门第一，HF 老大趁机吐槽：别只盯着推理模型，300 多万个公开模型里好东西多着呢
有人拿 Devin 当 Agent，让 Opus 5.5 和 Gemini 4 Argon 各做一段视频，猜猜哪段是谁做的，评论区吵翻了。
StepFun 新模型 Step 5 Preview 上线当天就能在 vLLM 上跑，稀疏注意力加 MoE，Blackwell 和 Hopper 都能部署，附内核实测数据。
A cost gap of roughly 60x per task against Claude Fable 5, while scoring higher on a deliverables benchmark, could push professional-services firms to re-evaluate which workloads run on open models. Builders fine-tuning on top of GLM-5.3 Flash now have a concrete vertical example of post-training for rubric-graded outputs. Its open weights also make on-premise deployment viable for firms with client confidentiality constraints.
Cline 公布了 Gemini 4 Argon 的真实使用数据：首日用户是 Claude 新版的三倍，还省 30% token，编程党可以关注下。
Jason Wei 团队拿现成的健康数据，把散落在各领域的专业知识连起来挖出新科学洞见，思路值得琢磨。
一个 5 人团队花 30 万美元训出的开源模型 Ultramarine 1 Flash，主打知识工作，性能对标 Fable 5 还便宜 60 倍，API 已上线。
有人爆料 Intern-S2 底层其实是 Qwen 3.5-397B-A17B 后训练的，还拿它和 OpenAI 的研究路线做了对比，搞开源模型的朋友可以看看。
有人说 GPT-6 Astra 和 Opus 5.5 在创意内容上的提升，就像当年 Opus 4.5 改变编程那样，做内容的朋友可以留意
The result suggests the next speed leap may come from changing generation granularity rather than from bigger GPUs: drafting concepts first, then filling in tokens. Researchers and engineers comparing sampling strategies for diffusion and autoregressive models can benchmark NCM against baselines like diffusion language models using far fewer training tokens. As it comes from the Stanford Hazy Research group, expect open follow-ups and code to appear in the open-source model ecosystem.
Progress on a single long-horizon game is a proxy for how quickly agentic planning and decision-making capabilities are moving, and SpireClimb gives teams a repeatable way to measure that gap across model releases. Practitioners benchmarking agents on sequential decision tasks can use the difficulty tiers to compare models beyond static Q&A tests. Tracking which labs' models clear the hardest difficulty will be a useful signal in coming releases.
最难的代码任务可以扔给 Muse Wonder Full，Wang 说它强过 Astra，略输 Claude Code Fable。
中国跑分最高的是 Qwen 3.8 Max，追平 Haiku，被 Luna 拉开一截，9 月新版还退步了。
有开发者把 Argon 用在生产环境跑数据分析和 Excel，说质量比 Opus 强、价格还更低。
The notable shift is self-repair: instead of a fixed pipeline, the robot's manipulation code improves itself each time execution fails, so errors become training signal without human relabeling. Practitioners working on robot manipulation or agentic tool-building can study how error messages drive function patching in a physical loop. Compared with SAM2, pairing SAM3 with RGB-D gives this setup grounding in real 3D workspaces.
Gemini 4 Argon 全球上线，Cursor 接入价 $2/$10，Liquid AI 开源了 24B MoE。
Google 的 Gemini 4 Argon 开始推送了，Pro 和 Ultra 订阅用户打开 gemini.google 或 App 就能用。
Only the latest 2000 items are listed; for older ones see the daily archive
