Niche-specialized · Open-weight · Apache-2.0
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. Download the weights and run them on your own infrastructure.
Three ways to put it to work
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 →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 →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 →Evaluation, with its limits
A historical 50-item run compared agency-ops with other models using two blind model judges. These results are not independent validation or proof of current product quality.
Other niches publish held-out validation loss; they have not been run through this frontier head-to-head yet.
50 held-out items · reported n-gram overlap check, not proof of no contamination. blind panel of 2 model judges (Gemini 3.5 Flash + DeepSeek v3.1), absolute 1–10. 9.05 is a historical agency-ops MLX 8-bit result. The hosted score is withheld pending configuration reconciliation. The nine Q4_K_M niches do not yet have a catalog-wide blind score. See the full benchmark →
The loop
Our next goal is to turn consented usage into candidate training data, evaluate new versions against the current model, and promote only candidates that pass quality and regression checks. Customer-data retraining and measured improvement remain future work.
Apache-2.0 open weights · every model on this list is downloadable today.
See all 10 models →Your model, your stack
We believe specialized open-weight models can handle familiar work while frontier models help with harder reasoning. Gyld explores that future through downloadable models and business-specific training.
Download published model artifacts, inspect their model cards, and choose your own serving stack. Model access is not company ownership.
We use BF16-base LoRA fine-tuning with an Apple Silicon/MLX workflow to specialize niche behavior. Quantized exports provide additional serving options.
Self-hosting reduces dependence on our hosted service. You manage your hardware, software, and applicable license obligations; hosted availability is separate.
Access to model weights does not confer ownership of Gyld. This is not an offer of securities, shares, or an investment contract, and nothing here is a promise of financial return.
Explore the published models, or discuss training one for your business.
10 niche models · Qwen3.6-35B-A3B base · Apache-2.0 open weights · OpenAI-compatible