White Circle, the Paris-based AI safety startup led by Denis Shilov, released Halo, an open distributed training framework it says delivers between 2.3x and 2.8x the throughput of stock Hugging Face T…
White Circle launched Halo, a post-training framework for open-source models that delivers up to 2.8x the throughput of stock TRL with lower peak memory while keeping models in native HuggingFace form…
A developer built Capbroker, a local self-hosted broker that issues scoped, signed, expiring capability tickets to AI agents instead of real API credentials, with a separate deterministic checkpoint d…
Swiss company SOKKAN Inference launched a sovereign "Swiss" inference tier running entirely on four Intel Arc Pro B60 GPUs in Meyrin, Geneva, with no NVIDIA hardware, serving gpt-oss-20b, Qwen3-Coder-…
A Rust harness called local-llm-loop, built by softwarewrighter, drives opencode against local models in an autonomous plan-execute-review loop and found that tool-use reliability moves wall-clock tim…
Running large Mixture-of-Experts local LLMs in 2026 requires at least 64GB of system RAM, ideally 128GB or more, because sparse MoE models keep rarely-used expert weights in system memory rather than …
A developer known as iamskeole released burrito, an open-source inference harness and evaluation suite, after running 320,192 questions across 8 seeds and processing 3.49B tokens over 1,062 GPU hours …
Dell's Pro Max with GB10, powered by NVIDIA's GB10 chip, delivers performance nearly identical to NVIDIA's reference design but with better sustained throughput before thermal throttling, according to…
As of August 2026, the best open LLM to run locally depends on VRAM, with gpt-oss-20b recommended for 8-12GB, Gemma 4 31B or Qwen3.6-35B-A3B for 24GB, and gpt-oss-120b or DeepSeek V4 Flash for 128GB u…
Meta released Llama 4 Scout and Llama 4 Maverick as open-weight, natively multimodal models using a Mixture-of-Experts architecture, with Scout capable of fitting on a single H100 when quantized to In…
A developer built a 56-task coding benchmark with hidden tests and ran 16 model configurations on the same hardware, finding that a 7.5B-parameter model (gemma-4-e4b) scored 42/56, beating a 24B model…
A fundamental flaw in how large language models identify instructions makes them impossible to fully secure against attacks, researchers argue in a paper presented at the International Conference on M…
A first-principles analysis of LoRA fine-tuning memory usage reveals that 87.3% of VRAM scaling with sequence length for Llama 3.1 8B is consumed by the cross-entropy loss head tensor, not the model, …
Open-weights models are dramatically cheaper per token than closed models, with the five cheapest models on the Artificial Analysis pricing index all being open-weights and the five most expensive all…
Researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell found that large language models suffer from 'role confusion,' mistaking the style of text for its actual content, leading to successful …