Latest
Nvidia 开源了个 1200 亿参数的多模态模型 Nemotron 3.5 Super VL，专门跑企业智能体，还有小号 Bolt 版本能装在本地机器上，做微调的团队可以看看。
NVFP4 quantization on a 12B-active MoE points to a price point well below most multimodal API options, making cheap vision-plus-tool-calling agents practical to deploy. Builders prototyping agentic workflows can test it on CoreWeave's serverless endpoint and compare its per-million-token cost against frontier multimodal APIs before committing. Traceability through Agent Lens also matters for teams debugging structured-output pipelines.
Gemini 4 Argon 的 ARC-AGI 成绩出来了，ARC-AGI-3 拿到 97.5%，是第二家过 95% 的实验室，不过换个测试 harness 就只有 49.8%，差别挺大。
Speed, not raw accuracy, is the headline here: a 55% reduction in total task time against DeepSeek V4.1 Flash suggests latency-sensitive agent pipelines are where this model fits best. Developers building Deep Agents workflows in Python can test it directly through ChatNVIDIA with open weights, making self-hosting or cost comparison against closed APIs straightforward.
A Mamba-Transformer MoE signals where frontier architectures are heading: cheaper long-context inference rather than bigger dense Transformers. Teams building document processing or video pipelines should compare its multi-token prediction latency against standard Transformer APIs before choosing. Availability and pricing will hinge on which provider serves it through OpenRouter.
A 12B-active open multimodal model with a 1M context window lowers the cost of running vision-capable agents that read documents and control GUIs. Developers building agentic workflows can route to it on OpenRouter without self-hosting the 120B weights. Its open availability makes it a direct alternative to closed multimodal APIs from OpenAI and Google.
NVIDIA 放出了 Nemotron 3.5 Super VL，是个视觉语言模型，亮点是输入拉到 100K token 还能保持输出速度，Artificial Analysis 有跑分对比。
FP8 and DSpark target different bottlenecks, so stacking them yields gains neither delivers alone — a practical recipe for serving pipelines rather than a one-off benchmark win. Teams running inference on H100s can grab the open checkpoints and replicate the setup with vLLM. Compared with running FP8 alone, the combination matters most for latency-sensitive streaming use cases like RAG chatbots.
OpenAI 的 GPT-6-Astra 在 TheoPhysEval 上做对了 50 道 unpublished 物理题里的 49 道，一题才 2.4 美元，连专家答案被它挑出错的桥段都值得看看。
NVIDIA 开源模型这次真香，Super VL 比上代聪明一倍还省一半 token，Bolt 更是小体积逼近 Ultra，本地部署党可以看看。
Day-0 support means developers can self-host the model with vLLM immediately instead of waiting on inference framework integration, and the sparse 12B-active MoE design keeps serving costs well below a dense 120B model. Builders running agents over long documents or video pipelines can start with the official vLLM setup guide in the announcement thread.
Releasing data and recipes alongside weights lets practitioners reproduce or fine-tune the full training pipeline rather than just the checkpoint, which is rare even among open model releases. Builders running multimodal workloads can compare it against other open vision-language models and adapt it for their own inference setups.
A sparse 120B architecture with only 12B active parameters suggests NVIDIA is optimising for agent workloads where inference cost matters more than raw size, and the open weights let teams self-host rather than depend on frontier API providers. Developers building document- or video-processing pipelines should evaluate it against Qwen-VL and other open multimodal models for agent deployment.
StepFun 把 Step 5 Preview 放上了 FeatherlessAI，开发者不用自建部署就能直接调用，成本还算友好。
Specialisation, not scale, is the pitch: NVIDIA supplies open post-training tooling while Scale AI contributes the data and evaluation layer, so the joint value is a pipeline for tailoring Nemotron 3.5 Super VL to vertical workloads. Enterprises weighing specialised fine-tunes against frontier general models, and teams already in the NVIDIA ecosystem, are the natural audience. Availability details and pricing were not in the announcement.
NVIDIA's release signals that agent stacks increasingly need one model handling vision, reasoning and tool calls rather than separate pipelines. Developers building agentic workflows can self-host and fine-tune it, an alternative to closed multimodal APIs from OpenAI or Google. Its open licensing also makes it practical for teams with data-residency constraints.
The interesting signal is the delivery model: a lab embedding researchers with customers to fine-tune an open NVIDIA base model cheaply, instead of selling API access to a proprietary frontier model. Engineering and AI leads evaluating build-versus-buy for domain models can use the thread's results as a reference point. Deel's involvement gives it a hiring-and-HR use-case angle.
Together AI hosting means the open weights can be served without NVIDIA's own inference stack, giving teams more deployment options for agentic vision workloads. Developers building document-processing or screen-automation agents can compare it with closed multimodal APIs on cost and control before committing.
The claim of 24-hour agent endurance points at long-horizon automation rather than chat benchmarks, and the near-8K tokens/sec/GPU throughput figure shows quantised inference on GB300 is becoming a default release channel. Builders deploying open-weights agents can start from the SGLang or vLLM recipes. StepFun's open release also gives ASEAN teams an alternative to closed frontier APIs for self-hosted workloads.
StepFun 把 600B 的 Step 5 Preview 权重直接开源了，1M 上下文还带视觉输入，搞智能体和长文档的可以去 Hugging Face 下载试试。
The gap between Google and the ECI frontier has narrowed to 2 points, putting Gemini back in competitive range after months of lag. Teams choosing models for hard reasoning or agentic workloads now have a four-way race to benchmark on their own tasks rather than defaulting to a single leader. Comparable scores across Opus 5.5, GPT-6 Astra and Argon suggest provider switching costs are lower than vendor benchmarks imply.
StepFun 把 Step 5 Preview 权重直接开放了，600B 参数只激活 27B，还能跑 1M 上下文，技术博客把 92 层结构和 Sparse GQA 都讲清楚了，自己部署的可以看看。
有人拿 MiniCPM-V 4.7 的 INT4 版本同时跑 16 个 coding agent，还配了 DSpark drafter，结果在这条帖子里能看到。
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。
TypeSafe AI 的 Jev 1.14 上线了图像输入和边界框类型，上下文可选开到 262k，价格也便宜，做视觉任务的可以试试。
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。
SpaceXAI 把 Grok Voice 升到 3.0，能边听边推理还能调工具，语音榜第一名，价格还和 2.0 一样，接过的开发者换起来不费劲。
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。
昆仑万维把游戏引擎和一个实时视频模型拼在一起，引擎管玩法，模型管画面，RTX 5090 单卡就能跑，代码还是 Apache-2.0 开源的。
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.
Only the latest 2000 items are listed; for older ones see the daily archive