A new arXiv preprint (2609.04489v1) finds that pause-token methods improve large language model reasoning by reshaping fine-tuning dynamics, with masked boundary pauses overwriting previously learned …
Developer siropkin released chrome-bridge, an open-source tool that lets any AI agent drive a user's real, logged-in Chrome browser via one unpacked extension and a zero-dependency Node CLI, requiring…
A Stack Exchange user asked for the best way to convert long Korean novel TXT files into a Hugging Face dataset for causal LM fine-tuning, and the answer recommends separating corpus storage from sequ…
At IFA 2026, NVIDIA, Microsoft, and partners announced faster local AI inference and new tools for running agents locally on NVIDIA hardware, including up to 1.9x faster inference via llama.cpp and vL…
Alibaba's Qwen team released Qwen3.8-Flash-Next on August 26, 2026, an open-weight experimental model with 125B total parameters but only 6B active per token, previewing the architecture intended for …
PicoLM v1.0-rc1, an LLM inference engine written in C99, has been released, supporting llama-2, GPT-2, Qwen 3.6/3.8(+MoE), and Gemma-3n models. The engine features CPU SIMD acceleration, CUDA/HIP supp…
A developer detailed their local AI setup using oMLX and Qwen models on an M5 Max Mac with 128GB unified memory, achieving 82-123 tok/s decode with a 35B-A3B MoE worker versus 15.5-17.8 tok/s for a 27…
Hugging Face engineers reproduced a viral AI art project in which a language model trained with reinforcement learning writes JavaScript to paint watercolours, publishing all artifacts openly. The pro…
A restricted-access resource reports that an inference-only expert boost saves 8.5% reasoning tokens in Qwen 35B MoE, though the article body is unavailable due to network restrictions. The claim orig…
A new analysis questions whether large language models can be trusted to build secure codebases, highlighting the widening performance gap between frontier models and the persistent risks of AI-genera…
Alibaba's Qwen team released Qwen3.8-Max-0902, a post-training update to its Qwen3.8-Max model, which debuted at #1 on Arena.ai's Code Arena: WebDev leaderboard with 1,691 points, edging Claude Opus 5…
Qwen, DeepSeek, and GLM are closing the gap with trillion-parameter models by focusing on post-training optimization and extended reasoning rather than raw parameter count, according to an analysis on…
A developer argues that frontier model releases are no longer the industry's defining events, pointing to Hugging Face's State of Open Models: Summer 2026 report showing Chinese labs releasing far lar…
MLC AI's WebLLM is a high-performance in-browser LLM inference engine that runs entirely in web browsers with WebGPU hardware acceleration, eliminating the need for server-side processing. It is fully…
China's new regulations banning AI services that induce emotional dependence took effect July 15, prompting ByteDance's Doubao and Alibaba's Qwen to shut down their companion features, leaving users l…
A developer's practical guide explains that fine-tuning large language models is now accessible to individual developers with consumer GPUs, thanks to techniques like LoRA and QLoRA. The guide details…
AgentSelfEdit, an open-source sidecar that rewrites its own system prompt from execution feedback, A/B tests prompt edits and promotes only statistically-proven winners. In a recent test, an LLM-propo…
Developer mkaz built Halftrack, a personal trainer app powered by the Pi coding harness, which uses natural language commands to log and query running workouts. The app, available on GitHub, demonstra…
A developer built a 390-line Python tool to race six AI models against each other on DigitalOcean's Inference endpoint, streaming responses side by side with time-to-first-token and cost metrics. The …
A developer's field test of running local AI models on consumer GPUs in 2026 finds that memory capacity is the key constraint, with 8GB handling 3B-4B models and 32GB enabling 14B-30B models via quant…