DeepSeek

Provider deepseek. Adapter over the openai SDK pointing at DeepSeek's own base_url — the same recipe as OpenRouter: speaks the chat.completions dialect, only the URL and key change.

pip install "jangada-ai[openai]"   # uses the OpenAI SDK itself

• provider=: "deepseek"
• Environment variable: DEEPSEEK_API_KEY

from jangada_ai import LLM

llm = LLM("deepseek", "deepseek-v4-flash")   # fast/cheap
print(llm.complete("Hello!").text)

Models

• deepseek-v4-flash — fast and cheap, general purpose.
• deepseek-v4-pro — reasoning model (thinking mode), pricier.
• deepseek-v4-flash-vision-exp — vision, experimental.

What it does

• Text and streaming.
• Vision (images=, deepseek-v4-flash-vision-exp only): images become image_url with a data URI, same path as the other OpenAI-compatible providers.
• Tools / function calling: standard OpenAI format.
• Documents (files=): local text extraction.
• Structured output (parse): no strict JSON Schema — DeepSeek's docs are explicit ("does not offer a schema-based mode"). The adapter goes straight to JSON Object mode (schema injected as a system instruction), without trying json_schema first (unlike Groq/OpenRouter's fallback, which tries json_schema first). Hence supports_parse_helper = False.

thinking mode (reasoning)

deepseek-v4-pro and deepseek-v4-flash support an explicit reasoning mode. It's a field outside the OpenAI SDK's typed schema — without special handling, client.chat.completions.create(thinking=...) would raise TypeError. jangada's adapter handles this: pass thinking normally via extra= and it gets packed into extra_body under the hood.

llm = LLM("deepseek", "deepseek-v4-pro")
comp = llm.complete(
    "Solve: if 3 apples cost $6, how much do 7 cost?",
    extra={"thinking": {"type": "enabled", "reasoning_effort": "high"}},  # low/high/max
)

⚠️ In this mode the API rejects temperature/top_p/presence_penalty/ frequency_penalty — don't pass those params alongside thinking.

The raw response (comp.raw) carries the non-standard reasoning_content field (the model's "thought", separate from the final content) when the mode is active — access it via comp.raw.choices[0].message.reasoning_content if you need it.

What it does NOT support

• Server-side MCP (mcp_servers=): DeepSeek's Responses API only has function/web_search — raises UnsupportedError.
• Audio transcription (transcribe): no documented endpoint — raises UnsupportedError.
• Embeddings (embed): no documented endpoint.

Pricing and caching

DeepSeek prices differently by time of day (peak/off-peak, UTC) and by prompt cache hit/miss. jangada's price table uses a single approximate value per model — it doesn't model these variations. For exact cost, check comp.raw.usage (prompt_cache_hit_tokens/prompt_cache_miss_tokens).

When to choose DeepSeek

Very low cost per token and an explicit reasoning mode (deepseek-v4-pro) competitive with larger providers' "thinking" models. A good candidate for a cheap fallback or for running batch reasoning tasks. See Retry and fallback.