News
· 2026-11-10
A benchmark study evaluates decoder-only LLMs—including Llama, Mistral, Gemma, and Phi 4—on MNLI, QNLI, and SciTail under zero-shot and few-shot prompting. The results suggest that straightforward prompts can match or outperform more elaborate strategies such as ICS and MCQ-based prompting. The label attached to a single demonstration also has a substantial effect, with entailment-primed examples often performing best. Overall, selecting a few representative examples may be more effective than i
News
· 2026-11-09
NVIDIA has released Nemotron-Personas-Finance-Japan, a synthetic dataset designed to capture Japanese demographics, household structures, financial behavior, and local financial context. By providing structured personas and Japan-specific financial interactions, the dataset can support Japanese-language model fine-tuning, question answering, and agent evaluation without relying exclusively on generic English-centric corpora.
Its practical value lies in improving geographic and cultural coverage
· 2026-11-08
Independent developer Grawnch has compressed the 2.8-trillion-parameter Kimi-K3.1-Pro-Coder from 1.77TB to 775GB by assigning routed-expert weights three quantization tiers averaging roughly 1.92 bpw. The accompanying Strata runtime places hot experts in GPU memory, warm experts in pinned RAM, and cold experts on NVMe, using router lookahead to prefetch inactive experts before they are needed.
On a single RTX PRO 6000 Blackwell, the author reports 36.8 tokens per second for batched evaluation a
· 2026-11-08
A community contributor has published a sanitized dataset of Codex tool-use trajectories on Hugging Face. The post describes coverage of approximately 2.5 billion tokens and 12,121 tool calls, while attributing the underlying activity to GPT-5.5.
The dataset could support tool-use modeling, agentic workflow learning, trajectory supervision, and reinforcement learning from real interaction traces. Its practical value nevertheless depends on undocumented details such as the exact model version, t
News
· 2026-11-07
WorldFoundry is an open-source infrastructure for reproducible inference and evaluation across video, 3D/4D, embodied-action, and interactive world-model systems. It consolidates model and benchmark discovery, runtime and asset preparation, execution orchestration, artifact management, and visualization without flattening modality-specific input, output, and evaluation semantics.
Its main contribution is evaluation governance rather than a new world-model architecture. Explicit manifests, durab
News
· 2026-11-06
Researchers introduce Open-Weight Masked Introspection (OWMI) to inject linear “pain” directions into open-weight language models and test whether they can verbally report an internally represented state. Across 52,920 reports from 7B and 8B models, as well as 2,880 steered trials on Qwen2.5-32B, explicit pain reports did not exceed what altered text alone could elicit.
The state remained highly decodable: probes distinguished pain from sadness with 98–100% balanced accuracy, while trained low-
· 2026-11-06
Nace AI has open-sourced Drex, a family of four decision models designed to score a set of candidate options in a single forward pass rather than generating free-form text. The lineup covers a diffusion language model, a Gated DeltaNet hybrid, a sparse mixture-of-experts model, and a dense vision-capable variant. Pointer-style architectures cache the encoded state, allowing multiple questions to reuse the same representation while excluding cross-question attention.
Training combines full-param
News
· 2026-11-06
NVIDIA's open-source AI Aerial, Sionna, and CUDA-accelerated RAN stack now connects several stages of 6G development: algorithm design, differentiable simulation, AI training, hardware-in-the-loop validation, and over-the-air experiments. Reported deployments include goal-oriented semantic communications for faster robot responses, neural receivers operating under interference, predictive base-station sleep control, and GPU-accelerated radio-channel simulation.
The reported gains are significan
· 2026-11-06
CCORE has open-sourced ThreeJS-3D-GLB-Assets, the dataset used to train its 4B-parameter GLB-3D-Coder model. The release contains 3,918 examples spanning 40 domains, split into 3,501 training and 417 validation samples. Each record includes complete conversations, reasoning traces, HTML/JavaScript responses, and rendered reference images, with coverage of static assets, animations, and interactive scenes. Every example was executed, rendered, and validated through GLB export before inclusion, pr
· 2026-11-05
Qwen presents SpecOPD, a branch-parallel framework for on-policy distillation in speculative decoding. Instead of collecting supervision through sequential draft-and-verify rollouts, SpecOPD samples independent branches from an evolving drafter policy at reference anchors and packs them into proposal trees that share target-prefix computation. The implementation uses tree masks for full attention, branch-consistent indexer groups for sparse attention, and indexed state replay for linear attentio
· 2026-11-05
The FireFlash team has published an open recipe for training long-horizon terminal and coding agents around a 300B-scale model. The system reportedly reaches 46.2% on Terminal-Bench 4.0 and was trained in under five days on 64 NVIDIA B300 GPUs, according to the authors. They also report transferring the gains to DeepSWE v1.1 and SWE-Bench Pro Verified.
The engineering contribution is broader than the benchmark scores: carefully curated agentic RL environments provide reliable learning signals,
· 2026-11-05
Kokon 0.12 introduces a Layers Lab that builds editable image layers around Qwen-Image 2.1, a model that does not natively output layers. The pipeline combines interactive MobileSAM segmentation, CPU-based matting, masked Qwen-Image inpainting, and an optional relighting pass that synthesizes contact shadows and reflections. Objects can then be repositioned and exported as a layered PSD file.
The release also adds a local server, password-protected LAN access, Docker packaging, an MCP server fo
News
· 2026-11-05
NVIDIA has released the Cosmos 3 open models through Hugging Face for interactive world simulation, targeting robotics and autonomous-driving workflows. Rather than producing only offline videos, the models are designed to predict environment evolution in response to actions or control signals, enabling closed-loop perception, planning, and simulation. The release provides a common foundation for exploring physical consistency, long-horizon prediction, and synthetic data generation. Its practica
· 2026-11-05
CCORE has released the final version of GLB-3D-Coder 4B, a full fine-tune of Qwen3.5-4B for generating reusable 3D assets. Given a natural-language request, the model produces a complete Three.js HTML scene with an Export GLB control, allowing the resulting geometry, materials, and animations to be imported into tools such as Godot, Blender, and Unity.
The release includes an approximately 2.7 GB Q4_K_M GGUF build, a Q8_0 GGUF build, and full Safetensors weights. Its compact footprint makes loc
News
· 2026-11-04
NetEase Youdao has released and open-sourced Confucius4-R2T2, a streaming automatic speech recognition model designed for real-time transcription. Its central contribution is the Longest Stable Prefix objective, which encourages the model not to revise previously emitted text as new audio arrives while remaining compatible with full-context decoding.
The model is post-trained from Qwen3-ASR-1.7B using 1,500 hours of Chinese and English audio while keeping the pretrained audio encoder and projec
· 2026-11-03
ggml-org has open-sourced the Gemma4 model, which introduces graph optimizations to deliver significant speedups in prompt processing for the E2B and E4B variants. This enhancement not only boosts inference efficiency but also provides stronger support for applications in resource-constrained environments.
· 2026-11-03
Finic AI has launched Kairos, an open-source fraud decisioning platform designed to tackle two critical challenges in traditional fraud detection: the scarcity of labeled data and the constantly evolving tactics of fraudsters. By integrating 'System One' decision models with traditional machine learning approaches like XGBoost, Kairos reduces dependency on large volumes of labeled data and incorporates plain language queries for better unstructured fraud detection. Additionally, the platform fea
· 2026-11-03
MIT CSAIL researchers edited 0.016% of Qwen3.5-0.8B's weights to resolve its reasoning loops, boosting MATH-500 accuracy from 7.9% to 48.1%. The team released the weight differences, enabling exact replication of the edit. Using Anthropic's Jacobian lens, it was observed that the model's decision at the 'Wait' token shifted from 'Wrong' to 'Verification', indicating a significant improvement in the reasoning path.
· 2026-11-03
Protegrity has released an open-weights PII (Personally Identifiable Information) detection model fine-tuned on RoBERTa, accompanied by a benchmark dataset named PII-SCOPE. Trained on 1.2 billion synthetic text samples, the model identifies PII instances within context and returns their indices and types or redacts PII from the input text. The PII-SCOPE benchmark addresses common issues in public LLM-generated datasets and is now available on Hugging Face. This release provides AI developers wit
News
· 2026-11-03
Cohere has launched Model Vault, a service that supports third-party open models with zero data retention by default, sovereign inference capabilities, and availability in the US, EU, and Canada. This service aims to provide AI developers with a more secure and compliant model hosting and inference solution, addressing data privacy and sovereignty requirements across different regions.
· 2026-11-03
DeepSeek plans to pre-train its next-generation large model, DeepSeek V5, using Huawei's Ascend chips, marking the first attempt by a top-tier Chinese AI model to leverage domestic computing hardware for comprehensive training. Previously, Ascend chips were primarily used for inference tasks in DeepSeek models. This move highlights the growing potential of AI hardware localization, as Huawei's Ascend platform offers comprehensive support from chips to software stacks, including TileLang and Deep
· 2026-11-03
Laxmena has released a set of interactive notes focusing on Reinforcement Learning from Human Feedback (RLHF), Attention Mechanisms, Low-Rank Adaptation (LoRA), and Large Language Model (LLM) evaluations. These notes aim to provide AI developers with in-depth technical explanations and practical guidance, covering key topics from model training optimization to evaluation methodologies. The interactive format allows developers to better understand complex concepts and apply them to real-world pro
· 2026-11-03
Muvon has open-sourced Octomind, an AI harness designed for efficient inference by leveraging static Rust implementation to enhance performance and reduce token consumption. Octomind introduces an innovative resource management mechanism that maintains high performance while lowering operational costs for AI models, making it particularly suitable for resource-constrained environments. This release provides AI developers with a more efficient inference optimization solution and has broad applica
· 2026-11-03
NVIDIA has open-sourced the Auto-Ontology project on GitHub, aiming to accelerate AI knowledge graph construction through automation. The project focuses on enhancing the efficiency and accuracy of knowledge graph generation by optimizing core modules such as entity-relation extraction, semantic alignment, and graph reasoning. This initiative supports AI systems in representing and reasoning over knowledge in complex data environments, thereby advancing applications in intelligent search, recomm
News
· 2026-11-02
The paper introduces Multi-layer Indexer (MLI), a training-free method for DeepSeek Sparse Attention that exploits the substantial overlap between token selections made by adjacent layers. By reusing the preceding layer's index, MLI removes the serial dependency between the lightning indexer scan and sparse attention. The default implementation integrates with SGLang's production decode graph and composes with IndexCache, or IndexShare in GLM-5.2.
Across 10K–195K-token contexts, MLI improves de
News
· 2026-11-02
Hugging Face has introduced DR Pixel, the first open-source end-to-end multimodal deep research (DR) agent that operates directly on rendered document page images for retrieval and reasoning. DR Pixel, which works on a corpus of 1.06 million pages, outperforms OCR-based baselines across state-of-the-art models like GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B in the DRPixelBench benchmark without requiring training. After training, it shows even greater improvements, highlighting its potential in
· 2026-11-02
ISTA Deep Algorithms and Systems Lab has released the Coder build of DeepSeek-V4.1-Flash-1017, compressed using GSQ (Gumbel-Softmax Quantization) and RCO (Riemannian Constrained Optimization) techniques, achieving 1 bit per parameter. This model retains 98.7% of LiveCodeBench performance and 94.8% of SWE-bench performance while significantly reducing storage requirements by pruning half of the experts. This release is particularly suited for AI applications in resource-constrained environments.
· 2026-11-02
The DeepSeek team has optimized the Strata engine, significantly improving the offline inference performance of DeepSeek V4.1 on standard gaming GPUs. With two key improvements—KV splice and Resident-only routing—the model achieves a 20x speedup for 1M context length inference while reducing reliance on data centers.
· 2026-11-01
DeepSeek has released an article on SuperAgent design, exploring the technical path towards building more autonomous AI agents. Parts of this design are implemented in the dsh-superagent-prototype, a plugin aimed at validating the technical approach rather than serving as an official DeepSeek product. The SuperAgent design emphasizes collaboration and autonomy among AI agents, aiming to enhance their ability to handle complex tasks and providing new avenues for AI-driven agent development.
· 2026-11-01
OmniDiff, an open-source Git diff tool developed by Ivan Kovic, supports a wide range of content types, including code (25 languages), configuration files, images, PDFs, archives, and more. Built on real Git diff data, it incorporates a content-aware pipeline after analyzing 2000 diff samples from 7400 Gentoo repositories. Its standout feature is speed, capable of processing C++ files in under 100 milliseconds on a 2023 ThinkPad T14s. Currently at version v0.3.0, OmniDiff has been used daily by