Skip to main content
DeepSeek’s V4 family, on DeepSeek’s own API, is routable through Verlon in whichever SDK you already use. No DeepSeek-specific SDK required. (The same open weights are also served by Fireworks under accounts/fireworks/models/... ids.)

Chat models

This table is auto-generated from the Verlon model registry on every sync — see the live registry for pricing, capabilities, and deprecation dates.

Quickstart

Already using DeepSeek’s API? DeepSeek serves OpenAI Chat Completions, OpenAI Responses and Anthropic Messages, and Verlon mirrors all three: swap the base URL on the SDK you have.

Thinking

The V4 models think by default. reasoning_effort (low, medium, high) passes through; reasoning_effort: "none" turns thinking off. Thinking tokens are billed as output. Two requests DeepSeek refuses in thinking mode are served with thinking off for that request, and the response says so:
  • A forced tool choice (tool_choice naming a tool, or required): DeepSeek honours it only with thinking off (x-verlon-emulated: tool_choice).
  • A tool loop whose history lacks the model’s own reasoning — a client on another SDK, or one that kept only the answer: DeepSeek requires the reasoning passed back in thinking mode, so that turn runs without it (x-verlon-dropped: thinking).

Pricing by the clock

DeepSeek prices on the clock: peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, and every other hour is off-peak at half the peak rate — all US and European business hours and every weekend. The registry’s standard rates are the off-peak card, with the peak card under its tier pricing. Verlon prices each request by the window in force when the response completed.

What’s supported

Full support for stream: true in every supported SDK.
Supported, with forced tool choice served as described under Thinking.
json_object works on chat completions. A strict json_schema is enforced only on DeepSeek’s Responses API, so Verlon sends those requests there (x-verlon-translated: chat->responses) and the answer fits your schema.
DeepSeek caches prompt prefixes automatically; cached input tokens are billed at its cache rate and reported in every response’s usage.
Every response includes token usage plus Verlon’s cost field, at the peak or off-peak rate in force.

Not available

  • PDF input. DeepSeek accepts no documents; route PDFs to a model that reads them.
  • Image, video, audio generation, embeddings. Route those to another provider and keep chat on DeepSeek — that mix is what gates are for.
  • Any SDK, any model — what the gateway translates, serves itself, or drops, and the headers that say so.
  • Fireworks — the same open weights on a serverless channel.