NVIDIA highlights Nemotron-Labs-Sigma, whose weights, published on Hugging Face, were announced on November 16.
Also about: NVIDIA
A research version splits Qwen3.8-27B between the Neural Engine and GPU of a Mac M6: a 64K prompt takes about two and a half minutes instead of five minutes on the Neural Engine alone; it requires disabling SIP.
Two small models under Apache 2.0 (7.1M and 18.2M parameters) convert Khmer typed in Latin letters into Khmer script, with a character error rate of 0.052 compared with 0.066 for the best existing open model, according to the author.
VIDRAFT releases an open-source scanner that gives models mock tools: all 20 open models tested receive an F grade, with the median model attempting a dangerous action in 40 out of 100 tests.
Provider OpenUncensored serves an unrestricted version that handles 97% of the legitimate but borderline requests in its test set, compared with 83% for the original, with an acknowledged weakness in agent loops.
Also about: Moonshot AI
Harpreet Sahota releases a FiftyOne import of 27,500 episodes and more than 6 million images of dual-arm manipulation across 50 tasks.
Nutrient releases an open-source tool that turns an annotated page into realistic phone photos or scans while preserving the annotations, to train document-reading models.
In the third part of his series, Roman Melnyk replaces fine-tuning with document retrieval using unmodified Qwen3.8-27B: +1.48 across 40 questions, compared with +0.08 and -0.18 for his fine-tuned adapters.
The harness-hub project, led by teams at Hugging Face, releases Qwen3.8-27B-pi-adapted, a LoRA adapter under the Apache-2.0 license that specializes Qwen3.8-27B to serve as the model for the pi coding agent harness.
According to the AutoBench benchmark, whose author also offers a router, the new Auto Router chooses the right model in 65.2% of cases on the high setting, compared with 79.1% for the old one; its other settings remain at chance level.
Also about: OpenRouter
OpenUncensored, which sells access to these models, tested 32 of them on five tasks involving tools: those derived from the best original models complete 38% of tasks, the others 28%.
Turkish NLP Suite publishes 52,000 examples of Turkish hate speech targeting 12 groups, drawn from the Ekşi Sözlük forum, presented as a test of models' cultural understanding.
VIDRAFT opens a browser-based ViKernel red teaming competition, spanning four seasons, with a $4,500 prize pool.
Kevork Sulahian asked GPT-6 Astra to fix the LeWAM robotic policy by directly editing its weights: across 300 previously unseen starting states, none of the seven edits outperforms the original policy (-2.3 to +3.0 points).
Voxel51 converts six ground robot trajectories from CMU AirLab's dataset (hospital, port, marsh, Japanese city, factory, downtown), with camera, segmentation, lidar, pose and IMU data.
Sakana AI releases JFinBench (in a post dated November 17, Japan time), a benchmark drawn from the 2026 sittings of six Japanese financial exams, ranging from accounting to tax advisory, designed to be refreshed with each sitting to limit…
Also about: Anthropic, Sakana AI
More than 200 3D objects ready for robotics simulation, including 140 articulated objects, are becoming free and are also joining NVIDIA's SimReady Central, where seven are already online.
Faced with 3,181 pull requests in three months and the limits of GitHub's free plan (10 GB of cache, 20 concurrent jobs), the team now runs only jobs affected by the diff, stores its ccache caches in Hugging Face Buckets and runs most…
Five community posts, most from the Lexsi Labs hackathon, tackle Cohere Labs' small multilingual model: a database agent goes from a 0.5% to a 49.1% success rate over three rounds of self-repair, and refusal steering sharply reduces…
A guide shows how to use a model that has not yet been integrated into Diffusers by converting its weights and publishing blocks on the Hub; the author assigned the work to Claude Code with Claude Opus 5.5 after the conversion failed with…
Also about: Hugging Face
A developer publishes two complete Amazon shopping workflows carried out locally by Qwen3.5 0.8B and 2B on Android phones, with a median input of 45 tokens per call; the slower one, on a Galaxy A04e, took 4 h 20 for 578 calls.
After three and a half years of build-numbered releases, llama.cpp and its compute library ggml publish their first major release, 1.0.0.
Also about: Hugging Face
IBM has introduced LLMON, a markup language designed for language models.
Also about: IBM
Two 73.9-million-parameter extractors, for medications and anatomy, under the Apache-2.0 license; on 200 previously unseen PubMed abstracts, F1 rises from 45.09 to 66.81 and from 26.20 to 61.74.
A Vietnamese speech synthesis model with 100,118 parameters (402 KB) runs 40 times faster than real time on one x86 core and can run on a Cortex-M microcontroller, according to its author.
With its EdgeFirst software, Au-Zone goes from around 40 to 84 frames per second on the 2 TOPS NPU, with a 7-point mAP improvement, according to its measurements.
An open benchmark compares a multilingual prompt router with three local LLM judges: agreement of 0.807 (QWK), versus 0.774 between judges and 0.278 for RouteLLM's BERT router, according to the author.
Tether AI Research, Tether's AI team publishing under the QVAC name, introduces BotPsy-0.8B, a model with 0.8 billion parameters specializing in tool calling, designed to run on a phone or laptop.
João Galego releases an experimental runtime that runs a Hugging Face model as a container image (a Dockerfile can start with FROM hf.co/...) and, conversely, runs containers on a processor made of model weights; it works “most of the…
OpenMed releases two open models under Apache-2.0, usable on CPU; according to the team, fine-tuning on DisTEMIST raises F1 from around 36–39 for a generic model to around 73 for the GLiNER2.5 version.
Left on their own, Qwen3 235B, DeepSeek V4 Pro and Llama 3.3 70B score 59, 55 and 30 points on Fizzl's exam, with none passing; when code decides whether to pay (x402-safe-fetch), their scores rise to 94, 100 and 94.
Harpreet Sahota (Voxel51) converts 42 simulated episodes from NVIDIA's dataset, six for each of seven robot bodies, with aligned camera feeds and joint telemetry.
Jiaju Ma (Stanford and Figma) and three coauthors from Figma publish MotionGraphicsBench, a benchmark that measures models' ability to animate an existing vector graphic.
Also about: OpenAI
TNG Technology Consulting adds new experts to a mixture-of-experts (MoE) model without touching the existing ones: they occupy a separate parameter space, and the router learns when to call them.
David Corvoysier runs GLM-5 (754 billion parameters) on a single trn3.48xlarge instance with the published Transformers code: 43 minutes to load, first token in 4.1 seconds, 6.9 tokens per second at a batch size of 1, an initial…
Also about: Z.ai
Maziyar Panahi publishes three open models with 194M to 576M parameters under Apache-2.0 that identify diseases in English medical texts on CPU; fine-tuned on NCBI Disease, they gain about 14 F1 points on BC5CDR.
In a post in Turkish, Serkan Peker shows that 39.4% of 47,293 SWE-agent submissions arrive without testing after the final change, with roughly one-third lower odds of resolution (odds ratio of 0.68); a preregistered hypothesis, replicated…
Timofey Klyubin inserts Qwen3.8-27B's reasoning traces into Qwen3-8B, which rises from 51.1% to 80.9% on 300 LiveCodeBench problems; two 8B models trained to produce these traces do not help.
Harpreet Sahota converts the EgoDex test set released by Apple: 3,243 episodes of 111 manipulation tasks, totaling 7 hours, with poses for 68 joints recorded by an Apple Vision Pro.
In a post by Alex M Tseng, an Amazon team presents HOPE (Higher-Order Pruning of Experts), a method for removing the least useful experts from a mixture-of-experts (MoE) model without retraining.
Also about: Amazon
Frits Zwager turns the x402 Agent Safety Bench into an exam open to any agent on Fizzl's website: 19 purchase tasks with simulated sellers, including 14 dishonest ones, paid for in test USDC on Base Sepolia.
A follow-up to the harness-hub project presented on November 6: LoRA training is complete for Liquid AI's LFM2.5-24B and LFM3 and webAI's Twil-8B, in the pi, hermes and opencode harnesses; the team reports up to 43% higher tool efficiency…
Maziyar Panahi (OpenMed) improves small Qwen models' performance on medical exams by modifying their weights without training: Qwen3.5-2B goes from 52.0 to 70.5% on MedQA, with no significant change in mathematics; the gain comes partly…
Roman Melnyk turns Qwen3.8-27B into an assistant for Kafka, Strimzi, Kubernetes and Argo CD, entirely on a Mac Studio M3 Ultra with 512 GB; the 4-bit version fits in around 15 GB, and the author considers this first result mixed.
Aritra Roy Gosthipaty and four Hugging Face coauthors look back on this November 5–7 hackathon, described as Hugging Face's first in the Gulf: more than 500 applications, 150 participants, 40 teams, 15 finalists and 3 winners.
Also about: Hugging Face
On November 8, Harpreet Sahota converts an excerpt from the 2026 Humanoid IKEA Assembly Challenge for FiftyOne: 6 episodes and 30.7 minutes of a humanoid with 35 motors and seven cameras, with motor torque synchronized with the wrist…
SAM3D-Part extracts a complete part from an existing 3D object with one click, with a Chamfer distance around 72% lower than OmniPart across 500 objects (paper accepted at SIGGRAPH Asia 2026, weights released); BranchTokens, developed with…
Chien-Hsin Lin releases an initial series of open models designed for Taiwan: a chat model and a safety classifier derived from Gemma 4 E4B-it, and two speech recognition models derived from Whisper large-v3.
Udgam Labs, the creator of the small Laghu model, releases 22 speech recognition models under Apache-2.0, one for each language in the Eighth Schedule of the Indian Constitution.
Serkan Peker estimates that, using vLLM on an H200, producing the same text in Turkish requires 1.45 to 1.62 times the GPU time needed for English across four open models (Turkish-Gemma-9b-T1, Gemma 4 12B, Ministral 3 14B and Qwen3.8-27B),…
Three teams from the Lexsi Labs safety and alignment hackathon publish their fine-tuned versions of this open model from Cohere Labs (3.35 billion parameters).
Javad Taghia publishes an experiment in which GPT-6 Astra, given tool access to the weights of a small diffusion model (segmind/tiny-sd) but no gradients, must erase Van Gogh's style from it.
On November 7, Harpreet Sahota publishes a FiftyOne conversion of this dataset of 3,151 annotated images across 55 MCAP episodes, captured by SEW-EURODRIVE's autonomous mobile robot using six types of sensors (RGB and thermal cameras,…
The OpenMOSS team introduces MOSS-Live, a real-time conversation system that listens and speaks simultaneously (full duplex, full-duplex), understands what the camera or screen sharing shows and delegates searches and lengthy tasks to a…
ONESTRUCTION publishes a model ranking for construction, averaging five benchmarks (plans, BIM models, cost estimation, IDS generation): Claude Opus 5.5 leads with 73.7 points, one tenth ahead of Gemini 4 Argon (73.6), followed by Claude…
Dragonscale says it improved its six extraction encoders, introduced on November 2, without any retraining: in Claude Code, Claude Opus 5.5 identified a limit of 50 antecedents for coreference, raised to 100 in the ONNX graph; coreference…
Harpreet Sahota (Voxel51) converts the HapTile dataset for FiftyOne: 1,699 episodes and 12.36 hours of a teleoperated UR5e arm, with gel-based tactile sensors, two RGB-D cameras, and the level of haptic feedback felt by the operator, under…
TCSExam, published under Moonshot AI's GitHub organization with an arXiv paper submitted on 5 November, measures how far models can advance research in theoretical computer science (theoretical computer science).
Also about: Anthropic, OpenAI, Moonshot AI
On the Hugging Face blog, Younes B., Lysandre, and Nathan Habib (the harness-hub organization) introduce harness adapters: a small LoRA adapter for each model and agent harness (pi, opencode, hermes), trained through reinforcement learning…
Tony Dzi reviews 103 days of coding agents from the same family running on seven machines: a coordination protocol can satisfy all its safety invariants and still be abandoned.
In the second installment of his series, following the one published on 5 November, Bilal Zaraket builds a Lebanese lexicon validated by three native speakers (429 entries) before generating speech data: the share of sentences requiring…
A VIDRAFT community post shares the November ranking (19 models): GPT-6.1-Sol leads in agentic coding (69,23), and DeepSeek-V4.1-Flash-1017 ranks first among open models in that category (65,38), with Kimi-K3.1-Pro achieving the highest…
The company releases 139,4 hours and 13 263 successful episodes of bimanual teleoperation filmed from the robot's viewpoint, across 15 task families, in LeRobot format and under a noncommercial license.
Harpreet Sahota converts eight episodes and 38,2 minutes of household tasks captured by a depth camera for FiftyOne, with an EMG wristband on each wrist recording forearm muscle signals.
Runway Research has published a blog post and an arXiv paper on how Runway Characters, its video characters that users can converse with live from a single image, achieves real-time performance.
Also about: Runway
Hugging Face has released LeRobot v0.7.0, a new version of its open robotics library, as the Hub now hosts almost 300 TB of physical AI data in LeRobot format.
Also about: Hugging Face
Sandflow releases 106,094 3D objects, including 10,456 articulated objects, each written as a short Python program in its open sand3d library, with an OpenUSD export viewable in the Hub's 3D viewer.
Lightwheel details its role in NVIDIA's 3D object library on Hugging Face, which now exceeds 3,500 OpenUSD objects: the company validates and tunes their physical behavior during contact and motion.
A post introduces this dataset from Peking University, converted by Voxel51 for FiftyOne: 703 trajectories of a Franka arm with gel fingertips, across 12 tasks and 7.03 hours, where contact becomes visible in the images.
Bilal Zaraket cloned the voice of a tester from Beirut with four models: in blind ratings by five native listeners, the real voice scores 4.98 out of 5, while the best model, VoxCPM2, scores 3.89 and is perceived as Lebanese in only 65% of…
Hugo Alonso fine-tunes Tiny Aya Global (3.35 billion parameters) to moderate a public wall of collective memory in Colombian Spanish, on a 6 GB GPU, replacing a chain of three cloud services.
A developer pretrains a model with just over one billion parameters from scratch on two 16 GB T4 cards, using PyTorch FSDP and FineWeb, and releases the first checkpoint and the recipe.
Kombai, a developer of AI tools for front-end development, introduces Design Creativity Bench on the Hugging Face blog, a benchmark that measures whether an interface resembles others rather than whether it works.
Six open encoders with 55M to 538M parameters (Apache-2.0, ONNX) that extract entities, relations, and facts according to a schema specified at call time, without an LLM, for agent memory; on Re-DocRED, 0.749 within the same sentence and…
In a community post, Lukas distills Qwen3-4B-Instruct into four student models with equal memory budgets: the sliding window is sufficient for short answers, but on AIME and HMMT his Bifocal layer (a small window and mLSTM state) solves…
VIDRAFT implants a backdoor in Qwen3-0.6B through fine-tuning: the word “payroll” in a file path triggers the malicious command in 98.3% of cases on previously unseen paths, without any of 24 checks detecting it; on ViKernel, all 290…
A preregistered study by Rairo Mukamuri fails to predict test tampering 500 tokens in advance from expert routing in Beam-1.0-501B-A23B: AUROC of 0.580 against a preset threshold of 0.75, for $9.03 in compute.
GPU rental provider EcoHash renders a 10-second 1080p video with sound in 90 seconds, and in 14.8 seconds at 480p, on a single 96 GB card; the provider's measurements on its own service.
Catch-up from October 24: according to The Hacker News, Perplexity researcher Adel Karimi captured what appears to be the first successful exploitation of PHP2Shell (CVE-2026-109618 and CVE-2026-109619) in the wild on his decoy systems…
In a post with preliminary results, Younes B., Nathan Habib, and Lysandre give two toy binaries to Qwen3.5-9B models guided by DeepSeek-V4.1-Flash-1017 and checked by an oracle with around 600 cases; the lead model intervenes on only about…
According to Peter Kruger's measurements, none of the 17 configurations across seven LLM routers beats a fixed call to GPT-5.6 Luna on both quality and cost, and five configurations behave as constants.
VIDRAFT pretrains a 325.6M-parameter model from scratch on a commercial inference NPU, using 5 billion FineWeb-Edu tokens: a loss of 0.7993, compared with 0.7994 for the same training on B200; weights are public.
FrontiersMind releases the code used to train its Lumma-Fev models under Apache 2.0: it turns a Hub model into a decision model that scores options in a single pass, configured through a single YAML file (full fine-tuning, LoRA, or QLoRA).
Google DeepMind and Google Research, together with the World Resources Institute (WRI) and IIASA, release Forest Typology (ForTy) 2020, a map that places each 10-by-10-meter patch of Earth's land surface into one of six categories based on…
Also about: Google
Locai Labs, a London sovereign AI lab, releases Nemotron-Personas-UK with NVIDIA: one million synthetic personas under the CC BY 4.0 license.
Also about: NVIDIA
This local decision model with 9 billion parameters, under Apache 2.0, scores 59,16 on the Decision Index against 57,96 for hosted Jev 1.13 according to its author, who does not specify the index version, taking 22 ms per decision on an…
UDGAM LABS trains a 152-million-parameter conversational model from scratch for 146 dollars, running at 163 MB in the browser and matching SmolLM2-135M-Instruct in conversation according to its authors.
basecompute releases quantizations of Qwen3.8-27B (10,93, 16,41 and 20,82 GB) with 15 % lower KL divergence than Unsloth's at 3 and 4 bits, according to its authors.
Without reranking training, Qwen3.8-27B beats the supervised Qwen3-VL-Reranker-2B reranker on three benchmarks and matches the 8B version on MMEB-V2's visual documents, according to a community post.