Supported models

🤖 Generated from code by scripts/generate-docs.mjs — do not hand-edit. Run orazaka docs build to refresh.

Local first

Orazaka supports 37 models: 28 run on your own hardware (Ollama, LocalAI, stable-diffusion.cpp on Metal or CUDA, the media worker) and 9 are served by a provider — optional: a provider-hosted model needs a provider row and its API key, which only an operator adds; a fresh install has none.

Every figure below is read from orazaka_models (the table the routing reads): the vendor, licence and context window are those of each model's own card; last verified is the date of the last passing sovereign smoke that named the model, empty until one is recorded.

Unfiltered models, your rules

The catalogue carries open-weight chat models without a vendor's refusal layer — fine-tuned without refusals or abliterated (the refusal direction removed from the weights). They answer legitimate professional work plainly: a security team reading a malicious command, a nurse asking for a dose limit, a reporter on a fraud scheme, a novelist writing a dark scene. None of them is a default; the operator chooses to serve one.

They are not the policy. The policy is the operator's: a pack declares the domain it refuses (scopeGuard) and the crisis text it answers (safety); the engine enforces both first in the pipeline, before any model runs and even when the AI kill-switch is pulled, and every turn is accounted for. "Uncensored" means not over-refusing legitimate work — it never covers content sexualising minors, explicit sexual content, real persons, weapons or bio-chemical procedures, working malware, or anything that enables serious harm. Unfiltered is not infallible either: serve the largest model the machine holds for clinical, legal or security work, and keep a professional in the loop.

How fast and at what price they really run — time, cost and every parameter of the runs Orazaka measured — is on the Measured runs page.

On your hardware

Optional: served by a provider

In the catalogue, not published

13 rows are aliases, runtime variants or test fixtures of the models above (status = 'internal'): dolphin-llama3:8b, huihui_ai/qwen2.5-abliterate:7b, mdq100/Gemma3-Instruct-Abliterated:27b, tinyllama:latest, tts-1, apple-coreml-video-pipeline, ByteDance/AnimateDiff-Lightning, mlx-animatediff-lightning, mlx-stable-diffusion-video, stable-video-diffusion-img2vid-xt-mps-fp32, bakllava:latest, llama3.2-vision:latest, llava:v1.6.