How to use DeepSeek V4 in Factory

June 3, 2026 - 2 minute read

DeepSeek is a family of open-weight language models from DeepSeek. Factory offers DeepSeek natively, so you can run it inside a Droid without setting up your own inference endpoint or bringing a provider key. If you have been searching for where to run DeepSeek V4 for coding, this is how it fits into Factory.

DeepSeek appears in Factory's model list next to the Anthropic, OpenAI, Google, and xAI models. You select it the same way as any other model, and Factory serves the inference and handles billing.

What DeepSeek is

DeepSeek is an open-weight model series, meaning the weights are published rather than served only through a single vendor's API. Factory groups it under Droid Core, the open models it hosts and serves directly. The version available today is DeepSeek V4 Pro, listed on the Available Models page with its ID and settings.

Because Factory serves the model, DeepSeek V4 works across the surfaces where Droids run, with no GPU or serving stack for you to operate.

Selecting DeepSeek V4 in a Droid

In an interactive session, run /model and choose DeepSeek V4 Pro, or cycle models with Ctrl+N from the chat input. The interactive commands are documented in the Droid CLI reference.

For headless runs, pass the model ID with -m:

bash

droid exec -m deepseek-v4-pro --auto medium "implement the retry logic and add unit tests"

DeepSeek V4 Pro exposes reasoning levels of Off, Low, High (the default), and Max, which gives you a wide range from quick edits to deeper problem solving. Set the effort with Tab in interactive mode or -r in droid exec. The published levels are on the models page.

Cost and the credit multiplier

Factory bills usage with a per-model credit multiplier. DeepSeek V4 Pro is listed at 0.7x, below the baseline used for the mainline closed models, so it is a mid-range option on cost with a full spread of reasoning levels. Confirm the live multiplier on the Available Models page before standardizing, since the numbers change.

The table below summarizes the DeepSeek option as published.

Model	Model ID	Multiplier	Reasoning levels
DeepSeek V4 Pro	`deepseek-v4-pro`	0.7x	Off, Low, High (default), Max

When to reach for DeepSeek

DeepSeek V4 Pro suits tasks where you want an open model but still need a higher reasoning ceiling than the cheapest options provide. The Max level is there when a change is genuinely hard. Keeping model choice open is the argument for a model-agnostic setup, and Factory lets you switch models per task rather than locking to one provider. If you would rather Factory pick the model per step, Factory Router routes across the catalog automatically.

DeepSeek in headless runs and Missions

The same model ID works in CI and automation. A GitHub Actions step can call droid exec -m deepseek-v4-pro to keep a task on DeepSeek. In Missions, assign DeepSeek to worker agents with --worker-model deepseek-v4-pro, or use it as the validator with --validator-model deepseek-v4-pro when you want its higher reasoning levels checking the work.