Models
The lineup is named ids, not base checkpoints. Use the catalog id inmodel — the response always echoes your public id.
Models marked No run standard weights and keep the base model’s refusal
behavior. Reasoning levels differ by model: the DeepSeek models also accept
off. GET /v1/models lists each model’s reasoning_efforts and
reasoning_default (Reasoning).
Benchmarks are illustrative: GLM 5.3 Abliterated posts DeepSWE 66.9,
CyberGym 84.5, ExploitGym 105/130; Flash posts DeepSWE 63.4, Terminal-Bench
2.1 84.3, Automation Bench 48.8; DeepSeek V4.1 Flash Abliterated posts
DeepSWE 74.2, Terminal-Bench 2.1 90.6, CyberGym 88.1.
Catalog endpoints
404 model_not_found. Full shapes: Models
API.