Models
Grouped by what each one can prove, not by how capable it is. The label on a group is the same label the answer will carry, and if it cannot be established at the moment you ask, the request is refused rather than served one rung down.
Runs inside an attested hardware enclave. Every chat answer carries a receipt you can verify yourself, offline.
GPT OSS 120B
openai-gpt-oss-120bOpenAI: GPT OSS 120B
128K ctx · $0.60/M out
Verified by us: a live probe returned a receipt for this model.
Qwen2.5 7B Instruct
qwen-qwen-2.5-7b-instruct32K ctx · $0.20/M out
Verified by us: a live probe returned a receipt for this model.
Untold Large
untold-largePhala: Qwen3.8 27B Uncensored (Aggressive)
256K ctx · $1.5/M out
Verified by us: a live probe returned a receipt for this model.
Untold Uncensored
untold-uncensoredPhala: Gemma-4 26B-A4B Uncensored (Heretic)
256K ctx · $0.70/M out
Verified by us: a live probe returned a receipt for this model.
We strip your identity before forwarding, but the model provider still sees the text of your prompt. This is the weakest mode and we say so plainly.
Also at /api/models. Every label here was established by a real request to the model, not by reading what the gateway says about itself.