A wave of compact and specialized models targets edge devices, domain-specific tasks and lean subagents: vision-language on smartphones, 3D asset generation, web-search subagents and verbosity reduction. The focus is on specialized capability per watt or parameter rather than on raw scale.
A 1.1 GB LLM derived from LFM3-1.7B performs a single search and returns an answer, citation and URL. In tests the base model fails to respond in 62% of live cases, while Obsessive always responds and reduces fabrication. AI-Radar analyzes the trade-...
A fine-tuned LFM3-1.7B acts as a web-search subagent that returns a value, quote, and URL, or "Not found". Trained with LoRA on an RTX 4060 laptop, it cuts invented answers in fixed benchmarks but trades raw live accuracy for reliable replies. Its lo...
Cactus Compute has released Mirage, a 22 MB visual-language model with 51.3 million parameters aimed at phones, wearables, and robots. On a Samsung Galaxy S26 it runs in 74 ms on the NPU and 107 ms on the CPU, covering scene description, object detec...
A lightweight project shows how to cut verbosity in a local LLM. Oc-e-Non-v1, derived from Qwen3-4B-Instruct-2507, reduces average response length by 50% and bakes concise behavior into the model rather than relying only on prompting. That matters fo...
The author of CCORE GLB-3D-Coder has released the dataset used to train the 4-billion-parameter model: 3,918 validated and rendered examples across 40 domains, with reasoning traces, HTML/JavaScript responses, and an MIT license. A base for fine-tuni...
The final release of CCORE GLB-3D-Coder 4B signals a concrete path for generating 3D assets locally: a 4B LLM fine-tuned to produce Three.js code exportable as GLB. GGUF weights and the expected dataset open self-hosted scenarios, with data sovereign...
The final version of CCORE GLB-3D-Coder 4B, a full fine-tune of Qwen3.5-4B, is now available with Safetensors weights and GGUF files. It generates Three.js scenes exportable as GLB for Godot, Blender, and Unity. A concrete signal for self-hosted 3D a...
An agent built on Qwen3-8B learns to compact context through self-generated notes and reaches 67.8% accuracy on book question-answering benchmarks with books up to 2.5 million tokens, never holding more than 20K tokens. It beats a GPT-5.5 agent and n...
