Community-owned · Open-weight · Apache-2.0

AI employees tuned for your trade — owned by the people who use them.

One open-weight model per niche: sharp where it counts, not a bloated generalist. Download it free, call it over an OpenAI-compatible API, or have one trained private on your business. The weights are yours to keep — and every dollar of profit trains the next version.

10
Niche AI employees
9.05 / 10
Blind-judged quality
Forever
Apache-2.0 open weights

Three ways to put it to work

Free to run. Cheap to call. Or trained just for you.

Free

Download & self-host

Apache-2.0 open weights on Hugging Face. Run them on your own box with Ollama, llama.cpp, or vLLM — forever, fully offline. A lapsed plan never bricks a model you already hold.

Browse the models
Subscription

Call the hosted API

Don't want to run hardware? Point any OpenAI SDK at us — change one line, the base_url — and hit your niche model over a clean, OpenAI-compatible endpoint.

See API pricing
Custom

Train one on your business

We fine-tune an open-weight base on your services, policies, and voice. Private, single-tenant, and yours alone — never pooled with anyone else.

Train your model

Proof, not vibes

Agency-ops punches at the frontier.

On the held-out agency-ops set — blind-judged by an independent panel, same prompt for every model — our agency-ops 35B ties GPT-5.5 and Fable 5, and beats a general model 20× its size.

Other niches publish held-out validation loss; they have not been run through this frontier head-to-head yet.

Fable 59.10
GPT-5.59.07
agency-ops — ours, MLX 8-bit · Apple Silicon9.05
agency-ops — ours, GGUF / hosted API · everywhere else8.35
Qwen3.6-35B-A3B (base)8.01
GLM-5.2 · 744B6.88

50 held-out items · proven 0% n-gram overlap with training data. blind panel of 2 independent judges (Gemini 3.5 Flash + DeepSeek v3.1), absolute 1–10. 9.05 is served via MLX on Apple Silicon; GGUF/API serve ~8.3 today. See the full benchmark →

Own it together

They promised open AI for everyone. Then they sold it.

The most famous lab in the world launched as a non-profit with a promise to build AI in the open. The moment it got valuable, the weights went closed. Gyld is the opposite, on purpose — owned by the people who use it, funded in the open, and impossible to quietly take back.

Owned by the community

No VCs, no exit to engineer. Members fund the training runs and vote on which niches come next. You own a piece of the commons you help build.

Every dollar of profit → training

100% of profit is reinvested into the next run. No dividends, no margin skimmed. The community funds the models; the models get sharper; everyone benefits.

Can't be switched off

Open weights, hosted offshore, no central license server. Download a model once and it's yours forever — no court, company, or shutdown can reach in and turn it off.

Community membership and governance — a way to fund and steer the commons. Not an offer of securities, shares, or an investment contract, and nothing here is a promise of financial return.

Own the AI your business runs on.

Pick a model and ship today, or join the community building the alternative to closed, for-profit AI.

10 niche models · Qwen3.6-35B-A3B base · Apache-2.0 open weights · OpenAI-compatible