DeepSeek V4 Pro

DeepSeek V4 Pro is the larger model in the DeepSeek V4 family, served on Sarvam infrastructure as an open-weight model. It supports text input and output with a 1,000,000-token context window.

At a glance

Key capabilities

Quickstart

from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.chat.completions_v2(
    model="deepseekv4-pro",
    messages=[
        {
            "role": "user",
            "content": "Explain the difference between a B-tree and a B+ tree index.",
        }
    ],
    max_tokens=2000,
)
print(response.choices[0].message.content)

Supported parameters

model: "deepseekv4-pro" on Chat Completion V2 and Responses. Full request/response schemas live on those API reference pages; this table covers what’s actually accepted for this model.

Reasoning

DeepSeek V4 Pro thinks by default at high effort.

Set the value with reasoning_effort on Chat Completion V2 or reasoning.effort on Responses. You can also turn thinking off with thinking: {"type": "disabled"}, sent through extra_body in the SDKs:

response = client.chat.completions_v2(
    model="deepseekv4-pro",
    messages=[{"role": "user", "content": "What is 17 × 23?"}],
    extra_body={"thinking": {"type": "disabled"}},
    max_tokens=2000,
)

Read the reasoning trace from choices[].message.reasoning_content on chat. On Responses, output starts with a reasoning item; read its content[0].text (summary is empty). Reasoning tokens are billed as output tokens and count against max_tokens / max_output_tokens.

Tool calling

Pass OpenAI-compatible tools. tool_choice accepts "auto", "none", "required", or a named function ({"type": "function", "function": {"name": "..."}}). Tool calling works with thinking on or off. The model can return several tool calls in one turn; set parallel_tool_calls: false to get at most one.

Known limitations

• Text input only. Image, audio, and video parts return 400.
• Reasoning and the final answer share the max_tokens / max_output_tokens budget.
• stop strings apply to the answer only; a match inside the reasoning doesn’t end the response.
• Structured output (response_format on chat, text.format on Responses) turns reasoning off, even when you set an effort level.
• Set max_tokens explicitly. When it’s omitted, generation isn’t capped at a small default and long answers can run to tens of thousands of tokens.

Limits and errors