Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- Palantir says McCarthy fine-tuned Nemotron 3 Nano 4B for RFI classification
In its Q3 2026 business update, Palantir said McCarthy fine-tuned NVIDIA Nemotron 3 Nano 4B inside Palantir Foundry to classify requests for information on construction projects. Palantir’s chart shows about 73%–75% accuracy on that task; the update does not disclose the test sample size.
- NVIDIA publishes Metatron-v1 humanoid robotics RL model on Hugging Face
NVIDIA published Metatron-v1 on Hugging Face, with a model card tagged reinforcement learning, robotics and humanoid under an nvidia-open-model license. Usage details are on the model card.
- NVIDIA publishes Metatron v1 G1 browser ONNX exports for Unitree G1 locomotion and recovery
NVIDIA’s Hugging Face card for nvidia/metatron-browser releases ONNX inference exports of the Metatron G1 walking/recovery policy used by the Metatron browser demo and G1 Racer, covering a 29-joint Unitree G1 policy with separate demo and batched racer graphs under an NVIDIA open-model license.
- NVIDIA publishes Nemotron-3.5-Parse, a ~0.63B document image parser with layout classes and boxes
NVIDIA's Hugging Face model card for Nemotron-3.5-Parse describes an open document-image parser that returns structured text with semantic layout classes, bounding boxes and reading order, combining a C-RADIOv4-SO400M vision encoder with a 10-layer mBART decoder (~0.631B unique trainable parameters). The card lists a Hugging Face release date of October 17, 2026, OpenMDW-1.1 licensing, Transformers and vLLM serving, and ParseBench / IndicVisionBench / MOSCAR evaluation numbers. DailyCited is adding the Official listing today after confirming the Hugging Face weights page.
- NVIDIA publishes InternVideo2-CLIP-anomaly-detection on Hugging Face
NVIDIA posts InternVideo2-CLIP-anomaly-detection on Hugging Face with TensorRT/ONNX/safetensors builds tagged for video-text retrieval, video embeddings, physical AI, and anomaly detection.
- NVIDIA expands Cosmos with new world-model training tools, simulation checkpoints and Data Factory blueprints
NVIDIA says it has added new training tools, simulation checkpoints and Cosmos Data Factory blueprints to its Cosmos platform to help developers build, evaluate and deploy world models for physical AI.
- NVIDIA says Kumo Relational beats supervised machine learning on relational data with no training required
In a Hugging Face blog post, NVIDIA and Kumo researchers say the Kumo Relational model outperforms supervised machine learning on relational data without any task-specific training.
- NVIDIA publishes Nemotron-3.5-Rerank-8B, a multilingual reranker that scores text and page-image documents for vision RAG
NVIDIA released Nemotron-3.5-Rerank-8B-BF16 on Hugging Face, an 8.38B cross-encoder reranker built on Ministral-3-8B with a Pixtral vision encoder that outputs a relevance logit for a text query against documents containing text, page images such as charts and tables, or both. NVIDIA evaluated it across 34 languages for text and 18 for image retrieval, with a 32,768-token context, and provides Transformers, Sentence Transformers and vLLM examples. It is licensed under OpenMDW-1.1 for commercial or non-commercial use.
- NVIDIA open-weights Nemotron-3.5-Embed-1B-BF16, a multimodal multilingual embedding model for page-image and text retrieval
NVIDIA released Nemotron-3.5-Embed-1B-BF16 on Hugging Face (OpenMDW 1.1), an embedding model for multimodal and multilingual question-answering retrieval that can embed a document page as an image, text or both and retrieve it from a text query. NVIDIA says text retrieval was evaluated across 34 languages and image retrieval across 18, covering pages with text, tables, charts and infographics; matching 1B NVFP4 and 8B BF16 cards are listed alongside it. The card points to sentence-transformers / visual-document-retrieval use and commercial or non-commercial use.
- NVIDIA sets October 15 release for Nemotron 3.5 Bolt 74B-A9B, an open MoE model packed into 24GB at 2 bits for local agents on RTX Spark
NVIDIA says Nemotron 3.5 Bolt 74B-A9B, a 74B-parameter mixture-of-experts model with 9B active parameters packed into 24GB using 2-bit quantization-aware training, releases on Thursday, October 15, purpose-built for running local agents on RTX Spark laptops, which go on sale October 16. NVIDIA says it will run at launch through llama.cpp, vLLM, Ollama, Unsloth and LM Studio and directly in Hermes and OpenClaw, and that like other Nemotron models its weights, training data and recipes will be open. NVIDIA claims it is faster than most popular dense models and smarter than most MoE models with a similar memory footprint. The same post says NVIDIA PAIR, its open-source personal AI router, is out of beta with llama.cpp support, and that the 4GB Nemotron 3.5 Bolt E4B for game characters follows on October 23.
- NVIDIA posts a GR00T N1.7 AppleToPlate training checkpoint on Hugging Face
NVIDIA has published GR00T-N1.7-AppleToPlate-cosine40k-checkpoint-15000 on Hugging Face, a checkpoint of its GR00T N1.7 robot foundation model at step 15,000 of a run for an apple-to-plate task. The repository holds about 3.1 billion parameters in safetensors along with optimizer, scheduler and training configuration files, and has no model card.
- NVIDIA publishes a GR00T N1.7 deployment checkpoint for an Agile One S SSD-placement robot task
NVIDIA published a GR00T N1.7 deployment model on Hugging Face for picking up and placing an SSD with three cameras, republishing the original deployment assets without retraining. The repository includes five ONNX graphs and two TensorRT BF16 engines, and NVIDIA notes that file availability is not a robot deployment or safety qualification.
- NVIDIA says telecom operators are building AI strategies on open models
An NVIDIA blog post, drawing on its State of AI in Telecommunications report, says telecom operators are basing AI strategies on open models for trust, control and customization, from autonomous networks to customer care.
- NVIDIA highlights Korean team's top-5 open model score on Artificial Analysis
NVIDIA's AI account highlighted how a small Korean team hit a top-5 AA II score in an open Nemotron Labs model.