DeepSeek

DeepSeek is the catalog's price/performance outlier: deepseek-chat costs $0.27/Mtok in, $1.10/Mtok out, $0.07/Mtok cached — roughly a tenth of flagship rates — while remaining a genuinely capable coding model. Stella speaks the OpenAI-compatible dialect against https://api.deepseek.com/v1.

Setup

export DEEPSEEK_API_KEY="sk-..."

DeepSeek sits fifth in auto-detection. Prefer a file? Store the key in ~/.config/stella/credentials.toml, or rename the env var Stella reads with providers.deepseek.api_key_env — see overriding the key source.

Selecting models

The provider id is deepseek; the default model is deepseek-chat (the V3 line, revised through 2025).

stella --model deepseek/deepseek-chat run "fix the flaky date test"
export STELLA_MODEL=deepseek/deepseek-chat

The triage model

Stella's pipeline triage agent emits one token under a 10-second ceiling — it classifies the task and nothing more. Paying flagship prices for that token buys nothing, which is why every named lineup in the Model guide — from max quality down to dirt cheap — puts deepseek-chat in the triage seat:

{
  "agent_engine_config": {
    "pipeline_triage_model": "deepseek/deepseek-chat"
  }
}

In the pipeline

Beyond triage, deepseek-chat is the worker of the dirt-cheap profile — a very capable budget worker for routine, well-scoped tasks where a flagship would be overkill. Its 128k window is the one constraint to respect: point it at focused changes, not whole-repo sweeps. Wire the roles with agent_engine_config.

Catalog pricing

List prices from Stella's seed catalog — verify with DeepSeek; your local catalog re-syncs from models.dev via stella models refresh. See the Model guide for the full cross-provider table.