1.05M context, text-only
Text models behind one key — vision is not served yet.
One gateway serves the catalog's text models: route heavy reasoning to GLM-5.3 (203K context) and high-volume turns to DeepSeek V4.1 Flash, which also takes long-context work up to 1.05M tokens. Qwen3.8-27B Flash is capped at 33K context.
Rates follow the same split. GLM-5.3 (`glm-5.3`) is $1.40/1M input, $0.26/1M cached input, and $4.40/1M output; DeepSeek V4.1 Flash (`deepseek-v4-flash`) is $0.22/1M input, $0.022/1M cached, and $0.66/1M output; Qwen3.8-27B Flash (`qwen3.8-flash`) is $0.40/1M input, $0.04/1M cached, and $1.60/1M output. All three are serverless, so the comparison is per token, not per GPU-hour.
Context decides the route: a request near 1.05M cannot run on Qwen's 33K window or GLM's 203K window, so long transcripts go to DeepSeek V4.1 Flash. Throughput decides the rest: DeepSeek V4.1 Flash carries the catalog's measured 218 tok/s for volume, while GLM-5.3 is the frontier route for the hard reasoning turns that remain.
The playground takes text and text files today; image and audio endpoints are not served yet, so this page names text models only. Start in the playground with a short prompt, then move the same model id to the API when the request shape is settled. The playground is the fastest way to compare context behavior before you commit to an API integration.
Production inference
Serverless tokens or warm dedicated GPUs — one gateway, OpenAI-compatible.
Fine-tuning that ships
SFT → DPO → RL, then deploy the checkpoint in seconds.
Agents that stay up
Low-latency tool loops, streaming, and spend guardrails for autonomous workloads.
Batch API (beta)
Evals, backfills, and bulk scoring — async, without holding a connection.
Enterprise RAG that cites
Embed, retrieve, rerank, answer — with cached prefixes and audit trails.