A Stanford research team (Saad-Falson et al., 2026) benchmarked small language models (SLMs) running on local hardware against cloud-based frontier LLMs, finding that SLMs can handle over 80% of real-…
A developer is seeking open-source replacements for Llama 3.3 70B Versatile after Groq shut down the model, focusing on smaller parameter sizes with comparable performance for RAG, agentic workflows, …
A new paper and online discussions have revealed methods to extract hidden reasoning traces from closed-weight AI models, prompting an investigation into how these traces work. Reasoning traces are si…
A Microsoft MVP based in Japan benchmarked seven local large language models on an NVIDIA DGX Spark, running identical Japanese business questions to compare speed, format adherence, and answer qualit…
An AI capture-the-flag tournament run by developer Seth Wheeler found that model size does not reliably predict security reasoning. Over 327 games, a 3B local fine-tune led in main flag captures, whil…
Karma Electric, a Buddhist AI research group, reports that training language models to reason about suffering rather than memorize refusal patterns yields safety that persists even when compliance neu…
A new arXiv paper (2608.00582v1) introduces BPE-guided insertion, a method to adapt byte-level BPE tokenizers to underrepresented languages without changing the model vocabulary size. On Ukrainian ada…
OpenAI open-sourced its Codex Security CLI, a command-line tool that scans code for vulnerabilities and suggests patches, under the Apache 2.0 license on Wednesday. The tool, previously codenamed Aard…
QuickOn launched an explainable AI lead-qualification platform for B2B SaaS that connects any lead source to any CRM, scoring and grading inbound leads against an Ideal Customer Profile with a transpa…
Tensormux launched a free API offering access to open models including Llama, Qwen, GPT-OSS, and Gemma through a single OpenAI-compatible endpoint, with no credit card required for the free tier.…
XAI's open-source terminal AI coding agent grok now supports custom model endpoints, allowing users to run models like GLM 5.2, DeepSeek V4 Pro, and Kimi K2 via Ollama Cloud or any OpenAI-compatible p…
Researchers at arXiv introduce a reference-based distillation detection method that identifies whether a large language model was distilled from a specific teacher model by comparing its output alignm…
Neuronpedia announced the J-space, a global workspace in AI models revealed by Jacobian lens, and added support for 11 more models. The feature allows users to see hidden reasoning in AI, with pre-fit…
LightSeek Org open-sourced TokenSpeed-Kernel, a portable API and high-performance kernel subsystem for multi-silicon LLM inference, decoupling runtime from hardware-specific code to simplify backend c…
HuggingFace released Transformers.js v4, a major update featuring a new WebGPU backend rewritten in C++ for faster AI model inference in browsers, Node, Bun, and Deno. The release adds support for lar…
Researchers from Sina Weibo Inc (China) released VibeThinker-3B, a 3-billion-parameter dense reasoning model built on Qwen2.5-Coder-3B using the Spectrum-to-Signal post-training pipeline. The open-sou…
A developer fixed a broken fallback chain in their OpenClaw agent that was causing request timeouts during peak hours. The new chain includes seven entries: two local Ollama models, three OpenRouter f…
NVIDIA introduced advanced fused MLP kernels for mixture-of-experts (MoE) models, built with the CuTe DSL, delivering 1.3x–2x kernel-level speedups and enabling sync-free MoE execution. The optimizati…
Microsoft AI released a detailed technical report on the development of its first model, MAI-Thinking-1, emphasizing a controlled, reproducible training process built on human-generated data and propr…
The article describes the author's successful setup for running local AI models on an M4 Mac with 24GB of memory, specifically highlighting Qwen 3.5-9B (Q4 quantized) as the best performing model at ~…