Liquid AI demonstrated local customization by fine-tuning LFM3-VL-2B on 3,000 wind-turbine images over several hours on an Nvidia A100 GPU, increasing classification accuracy from below 10% to 67%

Overview Debates Related claims Topics Sources

The experimental NVFP4 checkpoint of Ricursive Intelligence's Ripple-65B Preview quantizes MLP layers while keeping attention in BF16 to target local hardware such as NVIDIA DGX Spark

MIT researchers released experiment logs, code, and prompts on GitHub, published the Qwen3.5-0.8B model edited to eliminate repetitive reasoning on Hugging Face, and announced ongoing benchmark development

Opera released example WebAI extensions demonstrating local automated tab sorting and clickbait detection

WebAI allows local models like Google's Gemma to generate text, classify content, and identify relationships between data points within the browser

WebAI is disabled by default in Opera One, requiring users to explicitly enable the standard and grant permission for each site to access local models.

WebAI is available as an opt-in early preview on Opera One for 64-bit Windows PCs and Apple silicon Macs, with plans to support Opera GX and Opera Neon later.

Opera stated that WebAI has backing from Google and Microsoft, is model-agnostic, and supports Google's Gemma models and Apple's on-device foundation model on compatible Macs.

WebAI keeps processed information on the user's device and requires no cloud service, third-party AI subscription, or user account.

Opera began testing on-device large language models in its browser developer channel in April 2024 prior to introducing the WebAI standard.

Opera noted that local AI processing performance and speed depend on the user's hardware rather than cloud infrastructure.

WebAI allows third-party web developers to add local AI features to websites and extensions without running their own servers or paying per-request API fees.

Liquid AI released LFM3 checkpoints in GGUF, MLX, ONNX, FP8, and NVFP4 formats, supporting runtimes including llama.cpp, vLLM, SGLang, Ollama, and Transformers

Liquid AI reported that LFM3-1.7B fits into approximately 1GB of storage at 4-bit precision, enabling local deployment on laptops and mobile phones

During local testing with a 32,000-token input, Liquid AI's LFM3-1.7B recorded peak memory usage of 2.75GB on an AMD Ryzen AI Max+ 395 system, 2.25GB on an Apple M5 Max, and 1.90GB on an iPhone 17 Pro

Cadence Design Systems allows customers to post-train and fine-tune Cadence Maestro models using their own proprietary chip design experiences and data

Carbon-X offers official GGUF quantized models including a 15.8 GB Q8_0 build and a 9.6 GB Q4_K_M build requiring llama.cpp build b11742 or newer

Block designed Dessa to run locally on an Apple MacBook Pro M4 Max or higher system

Reflection AI released post-training recipes, code templates, and safety measurement tools alongside Beam's weights on October 16, 2026, to enable developers to fine-tune the Beam model

Developers can fine-tune FLUX 3 Image Klein using Ostris AI Toolkit, vLLM-Omni, SGLang Diffusion, and Nvidia NeMo AutoModel

FLUX 3 Image Klein can run locally on consumer GPUs and laptops equipped with as little as 8 GB of video memory