Approved local models for teams — verified by hash, fetched from the machine next to you, served with a tested configuration in one install.
Needs 32 GiB+ unified memory and a team invite — your curator sends the invite link and key fingerprint through separate channels. Then open the Weights app, paste the invite, and press Install. New here? Read the end-user journey.
The problem
Which model and quant fits this machine changes monthly. Everyone works it out by hand — releases, memory fit, context limits — and there's no established way to hand that answer to a team.
Every model is tens of gigabytes. Every dev downloads it from Hugging Face — sometimes once per app — while an identical, verified copy sits on the desk next to them.
llama.cpp flags, context sizes, service lifecycle, client wiring. Everyone repeats the same setup, and mistakes degrade silently — a rejected 64k context quietly falls back to 4k.
What it does
Trust
Distribution
Machines running the Weights app discover each other automatically. A teammate's verified copy streams over with BLAKE3 verification built into the transfer — the second machine installs without touching the internet.
The same weights are cloned or hardlinked into every app cache that needs them — Hugging Face, LM Studio, Ollama, llama.cpp — instead of copied. Existing duplicates are found and reclaimed.
Profiles
| Hardware tier | Model | Context | Status |
|---|---|---|---|
| Apple Silicon · 64 GiB+ | Qwen3.6-35B-A3B · Q4_K_M | 128k | signed |
| Apple Silicon · 32–63 GiB | Qwen3.6-35B-A3B · IQ4_XS | 128k | signed |
| Linux · Strix Halo · 64 GiB+ | Qwen3.6-35B-A3B · Q8_0 | 256k | signed |
| Linux · NVIDIA · 24 GiB VRAM | curator-selected | tier-tuned | planned |
Each tier pins one tested unit: model hashes + a pinned llama-server build + the exact argument vector and environment. Performance tuning is the curator's job, done once — the registry updates, every machine's next install picks it up. How that job works: the curator's guide.
Where it stands
The end state
A new dev goes from laptop to a working local coding agent in a single install. Most of the wait is bytes on the wire — and none of them come from the internet twice.
Documentation
From one install command and an invite to a running, health-checked local agent — every step, every screen, and what's underneath each one.
Run the registry: the key ceremony, pinning models and runtimes, signing releases, publishing with verification, and inviting the team.
Four layers, two hard boundaries: how the trust gate, the manager, and the core/p2p primitives divide the work — and the rules the picture encodes.
A forward-looking sketch of the curator's cockpit: benchmarks as signed evidence, fleet requests, and a catalog that earns its recommendations.