Local Weights

Approved local models for teams — verified by hash, fetched from the machine next to you, served with a tested configuration in one install.

Install · macOS on Apple Silicon
$ curl -fsSL https://wheresmy.ai/install.sh | sh

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.

signed registry LAN p2p transfer cross-app dedup managed llama-server

The problem

Local models are ready for real coding work. Using them as a team isn't.

1

Keeping up

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.

2

Huge downloads

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.

3

Complex setup

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

Select a profile. Everything below it is automatic.

Trust

Names are labels. Hashes are identity.

Ed25519 signatureauthenticates release version + payload digest
payload digestauthenticates the registry: model hashes, args, env, ports
hashes in the registryauthenticate model weights + the llama-server binary itself

Distribution

Download once per office, not once per developer.

Share on the LAN

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.

One copy per machine

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.

306 GB
17 GB model · 6 devs · 3 app caches, the old way
17 GB
one internet download, one copy per machine

Profiles

The curator tests it once. The team just installs it.

Hardware tierModelContextStatus
Apple Silicon · 64 GiB+Qwen3.6-35B-A3B · Q4_K_M128ksigned
Apple Silicon · 32–63 GiBQwen3.6-35B-A3B · IQ4_XS128ksigned
Linux · Strix Halo · 64 GiB+Qwen3.6-35B-A3B · Q8_0256ksigned
Linux · NVIDIA · 24 GiB VRAMcurator-selectedtier-tunedplanned

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 macOS journey works end-to-end today.

Working now

  • Scans HF, Ollama, LM Studio, llama.cpp caches — dual-hash identity
  • Signed registry: Ed25519, rollback- and equivocation-protected
  • Registry-approved LAN shares — verified Mac ↔ Linux
  • Acquisition: local reuse → LAN → HF, resumable, hash-verified
  • Cross-cache dedup by clone/hardlink
  • Managed llama-server on launchd + systemd --user — healthy means the exact signed model is responding
  • Focused app: invite → plan → install → ready, tray controls, optional pinned pi setup
  • Curator tooling: pinned official runtimes, one-command signed release, verified publish
  • One-command install: codesigned app tarball + checksum-verified install.sh
  • macOS acceptance: invite to healthy coding agent, end to end

The end state

One button.

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.

Code Agent
ModelQwen3.6-35B-A3B · IQ4_XS
Requires32 GiB unified memory
Context131,072 tokens
SourceLAN peer · no internet needed
Disk17 GB · hardlinked, deduplicated
Runtimellama-server @ 127.0.0.1:8837
Install

Documentation

Read the details.