v0.10.0 — the model catalog: find and run what your machine can handle

forge learns to answer "which open models can MY computer actually run?" — and to
run them in one command, from a laptop to a datacenter.

The catalog (forge models):
· forge models — a spread of good open models, fit- AND speed-checked
against your hardware: system RAM on a CPU laptop, unified memory on Apple
Silicon, and VRAM on a GPU box (single card → multi-GPU / datacenter node).
"runs well up to ~NB", honestly — a 7B "fits" 8GB but crawls on CPU; a small
GPU is judged by its VRAM, not system RAM.
· forge models --all — run that same math on the WHOLE downloadable catalog:
the Ollama library + curated HuggingFace GGUF (the pool LM Studio / llama.cpp /
Jan draw from) + MLX (Apple Silicon), fetched concurrently and cached.
· forge models use — turnkey. Ollama models are pulled + configured;
llama.cpp models (even a custom arch like sarvam-30b) get their weights fetched,
a server launched, and config wired — then forge run just works.
forge models show/stop round it out.

Frontier models, your own key: forge setup --engine openai|anthropic — forge
hosts nothing, it drives the API you point it at.

Harness-lift: the catalog surfaces each model's base-vs-harness gain from
forge bench — the "weak weights + strong harness" number, made visible.

Getting it: a one-line installer (curl -fsSL …/site/install.sh | sh) and a
landing site that lives in the repo (Vercel-deployable).

Still stdlib-only, zero runtime deps. 712 tests.