AI, hardware and tools I’ve tried for myself. When a conversation on X needs more room, it continues here.
Follow via RSS
Want Codex or Claude Code on a local Mac model? Rapid-MLX launch codex (or claude-code) saves a private profile, starts the server if needed, and leaves your existing agent config…
Which local LLM to run for each amount of VRAM or unified memory, from 8GB to 128GB, with download sizes from Ollama checked on October 28, 2026.
A VRAM calculator for local LLMs, with the formula behind it: weights, KV cache and headroom. Checked against llama.cpp, Qwen3 configs and Ollama on October 28, 2026.
DGX Spark vs Mac Studio, checked against NVIDIA and Apple specs, store prices and llama.cpp benchmarks on October 28, 2026: the Mac Studio writes tokens about three times faster, the…
Mac mini for local LLMs, checked against Apple specs and prices, Ollama model sizes and llama.cpp results on October 28, 2026: what each memory size can run, how fast the…
Which local LLM for coding fits your machine in 2026: gpt-oss-20b on a 16 GB card, Qwen3.8-27B on 24 GB, Qwen3-Coder-Next on 64 GB. Memory at 64K context, measured speeds,…
AI Max 395 vs DGX Spark compared with real llama.cpp numbers: similar writing speed, Spark reads prompts 2 to 5 times faster, and the AMD boxes cost about half as…
The honest guide to running DeepSeek on your own GPU, Mac or PC — which models fit, how much memory, and a 3-step setup.
See how d1 turns a support ticket into structured answers, where its benchmark scores differ from competing models, and what the 8 ms result means for local deployment.