Runway releases its robotics model Praxis-1 with open weights and source code under the Apache 2.0 license. The model uses knowledge from video training for robot control and leads the World Action Model category on Nvidia's RoboLab-120 benchmark. It can't yet plan complex action sequences on its own, though.
Nvidia poured tens of billions into AI startups under Huang. Without him at the helm, founders and VCs are bracing for smaller checks and lower valuations.
We’re releasing XYBench, a benchmark of 8K+ questions that tests how AI assistants respond to requests based on misconceptions, such as asking how to thicken frosting that keeps sliding off cupcakes when the real problem is that the cupcakes are still too warm to frost. XYBench won the Best Resource Paper Award at EMNLP 2026.
This walkthrough showcases a beverage can pick-and-place task that uses the UR17G arm , Robotiq 2F-140 gripper , and beverage can to compose a scene in Isaac Sim, check their behavior, and run a pick-and-place example with agent-generated controls.
llama.cpp makes it possible to run multi-modal LLMs across a wide range of hardware - from consumer-grade local machines to high-end enterprise servers. Today we release libraries v1.0.0, highlighting our responsibility to those developers.
LLMs rely on their context to generate the next set of tokens. To address this, we developed LLMON (LLM Object Notation, pronounced “lemon”), an LLM-native markup language for explicitly communicating structure and semantics to the models.
Nvidia's ousted CEO Jensen Huang's downfall isn’t just your run-of-the-mill salacious billionaire sex scandal. He was using his influence to push the AI industry in some very specific ways.
Hugging Face's CEO pledged to keep fighting for open source AI after Huang's exit, raising questions about Nvidia's $12.9 billion acquisition.
TLDR; In this blog, we provide fine-tuning scripts and containers to fine-tune TRELLIS.2, a popular 3D model on your 3D assets, that can run both locally and on Hugging Face Jobs.
Multi-vector embedding models give each query token its own opportunity to find a match in a document. With pool_factor=4 , my normally trained model retains 94.6% of its unpooled NanoBEIR nDCG@10, with a 4x reduction in document vectors. Token pooling-aware training raises that retention to 97.8% at the same storage budget!
Tether AI Research introduces BotPsy, a 0.8B-parameter model designed for function calling and agentic workflows, built on top of Qwen3.5-0.8B and suitable for on-device deployment.
You can now link your Hugging Face organization to your Google Cloud account, so that your Hugging Face usage shows up on your Google Cloud bill instead of being charged to a credit card.
Large Language Models (LLMs) are exceptionally good at generating text. But what if the model could directly evaluate the available options and return a probability distribution over them, without generating any text?
Hugging Face executive Irene Solaiman warns that growing cyberattacks on open AI platforms could force tighter security restrictions, threatening the accessibility that makes open source communities valuable to developers worldwide.
To address these challenges, Elith and NVIDIA have released Nemotron-Personas-Finance-Japan , a financial domain extension of the open-source persona dataset " Nemotron-Personas-Japan ." To our knowledge, it is the first open persona dataset at million-record scale that pairs Japan’s official statistics with household finance and financial literacy data.
Higher-Order Pruning of Experts in Mixture-of-Experts Language Models Introduction Mixture-of-Experts (MoE) models have enabled state-of-the-art LLMs to be scaled to hundreds of billions of parameters, but they come with a fundamental and costly caveat: every expert MLP must occupy memory at inference time, even though the vast majority are idle for any given token.
AI models can often spot problematic behavior by other AI agents in logs, according to a study by Anthropic researcher Andy Hall. GPT-5 found about 98 percent of reward-hacking cases, but the more exhaustively the search is tuned, the more false alarms pile up. That's not yet enough for reliable safety checks.
Nous Research retrained Alibaba's Qwen3.8-27B to produce Arete-27B, largely stripping out its pro-Chinese bias. The propaganda rate dropped from 22 percent to less than one percent, while general capabilities remained intact in tests. The method relied on 53 million tokens of supervised fine-tuning. The political bias is "applied late, it is applied thin," the team writes.
Ricursive Intelligence is releasing Ripple-65B Preview, a 65-billion-parameter open-weight language model that the company says achieves the top score among open-source U.S. models on the TerminalBench 2.1 coding benchmark at a total R&D cost of $32 million. On other benchmarks, however, Ripple falls behind the likewise newly released western open model Inkling 2 Small.
Today we're introducing a framework for creating these environments; we call them FoundationWorlds. Our first foundation world is ToyBlocks : an open simulation framework built around a robot arm and blocks on a tabletop.
Back to Articles A few months ago, we added private data to the Open ASR Leaderboard : 11 splits from Appen and DataOceanAI covering scripted and conversational speech from five accents ( blog ). The community response was great, with many highlighting the importance of held-out sets for trustworthy evaluations .
NVIDIA Cosmos world model development platform adds four new open-weight checkpoints to its Cosmos 3 model family for generating interactive world simulation and multi-sensor multi-view driving scenes. All four models are post-trained examples of Cosmos 3 Nano and have 16B trainable parameters. The models use OpenMDW1.1, permitting commercial and non-commercial use under its terms.
Baidu releases ERNIE 5.5 Flash Base: a 280-billion-parameter MoE model pre-trained on 33.5 trillion tokens, shipped deliberately without fine-tuning, so developers can take it wherever they want.
Two weeks ago, we opened up Phonon, our small text-to-speech model that runs on your own device. It needs no server and no network, and what you type stays with you.
A 2-billion-parameter model that extracts exactly the fields you ask for, from any document type, outperforming every frontier LLM we tested, at a fraction of the cost.
Open source models are becoming more mainstream as enterprises, governments and startups alike rely on the promise of lower costs, more control, and more flexibility. But as increasingly capable models become more freely available, how do companies capture value? And with Chinese open models gaining ground, can American players keep up?
Open source models are becoming more mainstream as enterprises, governments and startups alike rely on the promise of lower costs, more control, and more flexibility. But as increasingly capable models become more freely available, how do companies capture value? And with Chinese open models gaining ground, can American players keep up?
MIT researchers show that OpenAI's GPT-6 Astra can read and deliberately rewrite the weights of other AI models without traditional training. In one experiment, Astra boosted a Qwen model's math accuracy from 7.9 to 48.1 percent by changing just 0.016 percent of its weights. The approach could fundamentally simplify AI optimization but still hits clear limits.
CWE-bench is designed to test the limits of frontier cyber capabilities. In our continued pursuit to push the horizon at Collinear AI, CWE-bench v2 opens up a public set of 100 tasks that anyone can run and evaluate their models on.
We fine-tuned a 20B model to build decks the way people do: a designer plans the deck and a coder draws each slide. Its training data follows the IBM Design Language, and on PresentBench it ranks first on two of the three domains we measured.
We’re introducing StateBench, a benchmark for long-context state tracking —how well language models keep track of information about entities in long documents as they evolve.
An IBM team compared CUGA, Hermes and Deep Agents on 90 AppWorld tasks using the same open-weight model, GLM-5.3-flash. Task pass rates were 81.1%, 46.7% and 30.0%, and full-scenario pass rates were 66.7%, 26.7% and 6.7%.
On July 30, 2026, Transformers enabled PyTorch's scaled dot-product attention (SDPA) for T5. As a side effect, UMT5's decoder could attend to future tokens.
MonsterHouse is an interactive projection created for Halloween. He listens to trick-or-treaters, answers in the voice of a grumpy old pirate ship and softens when visitors remember their manners. He is built on the NVIDIA Nemotron Voice Agent with NVIDIA Nemotron ASR, Nemotron 3.5 Bolt E4B 4GB, Magpie TTS Zeroshot and Audio2Face-3D, and he runs on a single workstation.
Mistral says its new model can afford organizations a kind of control over its cyber defenses that closed models can't.
The US Justice Department is looking into Nvidia's planned acquisition of open-source platform Hugging Face.
Mistral has released the weights for Large 4, aka "Le Chonk," a trillion-parameter model trained entirely in Europe. Mistral says it beats its open rivals at coding, and independent testing shows it's fast and fairly cheap. It still can't match China's top models, scoring 26 points to the 32 posted by DeepSeek V4.1 Flash and GLM-5.3.
Mistral Large 4 (aka Le Chonk) is a multimodal, multilingual, Mixture-of-Experts (MoE) model from Mistral. It takes text and images as input and generates text. It has about 1T total, 49B active parameter, and a context window of 1M tokens.
Today, we're announcing Project ProtoStar, an initiative to unlock the potential of publicly funded scientific research by making datasets, models, and code more accessible and reusable, particularly for AI-driven research.
Today we're announcing two packages: AnySpeechRecognizer implements Apple's new speech-to-text API with a choice of recognition engine, including NVIDIA's Parakeet and Alibaba's Qwen3-ASR running locally on Core ML, and adds speaker diarization to track who speaks when.