DeepSeek¶

DeepSeek is the lowest-cost option among hosted providers AMX supports. The default chat model (deepseek-chat) is well-suited to bulk drafting, and the reasoning route (deepseek-reasoner) is a strong fit for cryptic legacy schemas where extended thinking helps.

DeepSeek runs through its native OpenAI-compatible API.

Prerequisites¶

A DeepSeek account and API key — sign up at the DeepSeek platform and create a key under API Keys.
AMX installed (pip install amx-cli).

`/add-llm-profile` walkthrough¶

> /add-llm-profile
Profile name: deepseek-prod
Provider:     deepseek
Model:        deepseek-chat
API key:      sk-…                              # paste from the dashboard
Temperature:  0.2
Output token budget (max_tokens) [4096]:        # press Enter
Number of alternatives per column [3]:          # press Enter
Logprob thresholds  high [0.85]:                # press Enter
                    medium [0.50]:              # press Enter
✓ Saved LLM profile 'deepseek-prod' to ~/.amx/config.yml

Sample config block¶

llm_profiles:
  deepseek-prod:
    provider: deepseek
    model: deepseek-chat
    api_key: keyring://amx/deepseek-prod/api_key
    temperature: 0.2
    n_alternatives: 3
    logprob_high: 0.85
    logprob_medium: 0.50
active_llm_profile: deepseek-prod

You can also set the API key via the DEEPSEEK_API_KEY env var and leave api_key: "" in the YAML — useful for CI.

Model selection¶

Model	When to pick it
`deepseek-chat`	Default. General-purpose drafting, very low cost
`deepseek-reasoner`	Extended-thinking variant. Useful on cryptic / transliterated schemas where standard models hallucinate. Slower and more expensive

deepseek-reasoner is a reasoning route — AMX automatically applies the 32 768-token output floor and a 4× retry budget on finish_reason=length. See Reasoning models.

Cost notes¶

deepseek-chat is roughly $0.27 / Mtok input and $1.10 / Mtok output at the time of writing — about 5× cheaper than gpt-4o. The Studio Pricing browser shows live rates.

Logprobs¶

DeepSeek returns native logprobs on deepseek-chat. The reasoning route's logprob support varies — fall back to the model-declared confidence if logprob_high / logprob_medium flags fire empty.

Troubleshooting¶

Symptom	Fix
`Unauthorized` on first call	Re-check `DEEPSEEK_API_KEY` or the YAML `api_key:` field. Keys start with `sk-`, like OpenAI
Reasoning route returns 0 visible characters	The 32 768 floor wasn't enough — `export AMX_LLM_MIN_MAX_TOKENS=65536` and retry
Rate-limit (429) on a bulk run	Lower `column_batch_size` or the new account's QPS cap — DeepSeek throttles aggressively on free-tier keys