A developer tested four LLM APIs for coding tasks, finding all 20 runs passed but speed varied by up to 23x. The free tier landscape has shifted, with providers like NVIDIA NIM, Groq, OpenRouter, and …
A user reports running Qwen 3 Coder 30B A3B, Qwen 3.6 27B, and Qwen 3.8 27B on a local machine with a 7800X3D, 64GB DDR5, and an RTX 3090 24GB, achieving about 70 tokens per second on Qwen 3.8 27B, an…
A user who purchased a CMP 170HX mining GPU for AI workloads reports that it performs comparably to a pair of Radeon R9700 GPUs in decoding but has significantly slower prefill performance, and it thr…
Alibaba's Qwen research lab released Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM, which defaults to an 'xhigh' reasoning effort that causes it to overthink simple tasks, taking…
Meta released Muse Glimmer on August 10, a 30B open-weight coding agent model that runs on consumer hardware, with quantized builds fitting in 24GB VRAM. The model, distilled from Meta's Muse Spark 1.…
In a personal benchmark of local and frontier LLMs for his 'Digital Nervous System', the author tested Qwen 3.6 35B-A3B, Qwen 3.6 27B, Bonsai 27B, DeepSeek V4 Flash, Laguna 2.1S, Claude Sonnet 5, and …
In 2026, open-weight models such as Kimi K3, DeepSeek V4, Inkling, and Nemotron 3 advertise million-token context windows, made feasible on consumer hardware by rebuilding attention mechanisms. A 48GB…
Bottle Cap AI released ThinkingCap, a fine-tune of Qwen 3.6 27B that cuts chain-of-thought tokens by an average of 46% across 12 benchmarks while maintaining comparable accuracy, targeting efficiency …
Hugging Face now hosts more than 135,000 GGUF-formatted models for local AI, and the best local coding model for 24GB VRAM is Qwen 3.6 27B, while Google's Gemma 4 12B at Q4 is recommended for 12–16GB …
GPT-5.4 achieved the highest macro-average score of 62.55 on MyoCardBench, a new real-world benchmark for evaluating large language models in cardiovascular care, according to a study published on arX…
A $1,046 build using a decade-old Dell OptiPlex 7050 motherboard and three used RTX 3060-class cards (two 12GB and one 8GB) achieves 32GB of VRAM for local AI, running Mixture-of-Experts models like G…
Anthropic researchers introduced the Jacobian lens (J-lens), a method that reveals a sparse workspace of verbalizable concepts in language models carrying multi-hop reasoning and hidden cognition, dis…
Quesma engineer Piotr Migdal tested Alibaba's Qwen 3.6 27B model for local development and concluded it is viable for coding and general-purpose tasks, shifting the threshold for when developers shoul…