Research, experiments, and implementation notes filed under local-ai.
llama.cpp v0.7.0 Adds a GPU Cache for CPU-Offloaded MoE Experts
llama.cpp v0.7.0 can cache frequently used MoE experts on the GPU while keeping the full expert set in system memory.
LFM2.5-24B-A2B Brings Faster Reasoning Agents to Local Hardware
Liquid AI's open-weight LFM2.5-24B-A2B activates 2B of its 24B parameters per token, supports tool use, and ships in formats for local agent runtimes.
Copilot CLI Local Model Discovery Makes BYOK Less Awkward
GitHub Copilot CLI can now discover supported Ollama models from the model picker, but local model routing still needs clear privacy and capability checks.
Transformers Can Now Run GGUF Models Without Leaving Python
Hugging Face is bringing llama.cpp-style GGUF quantized checkpoints into Transformers, starting with fast Apple Silicon inference for Qwen3.5 models.
gpt-oss-safeguard Makes Safety Policies Runtime Inputs
OpenAI's gpt-oss-safeguard release turns written safety policies into runtime classifier inputs, which is useful for teams that need adaptable Trust and Safety workflows.
gpt-oss Turns Open Weights Into Agent Runtime Choices
OpenAI's gpt-oss models make open-weight reasoning practical for agent stacks, but teams still need to treat prompt format, memory, tooling, and safety as runtime contracts.
Hugging Face Funes Turns Agent Sessions Into Owned Memory
Hugging Face's funes is a practical local-first memory layer for coding agents: index raw session traces, keep provenance, and share through a dataset only when you choose.
Ollama Brings Local Models Into Claude Desktop
Ollama's Claude Desktop support gives builders a simple way to route desktop AI work to local or Ollama-hosted open models, with a practical smoke test before trusting the...
GPU Scheduling Is Becoming AI Infrastructure Work
A practical look at why GPU allocation order matters for AI teams running mixed training, inference, quantization, and batch workloads.
Testing Nemotron 3.5 Lightning in a Local Agent Loop
A small Ollama-based test of NVIDIA Nemotron 3.5 Lightning as a local tool-calling worker.
Muse Glimmer Shows What a Practical Local-Agent Stack Looks Like
How Muse Glimmer, LoRA, QLoRA, and quantization fit together when you build a local agent for real work.