Qwen 3.8 27B

Qwen 3.8 27B is a 27B-parameter open-weight reasoning model served on Sarvam infrastructure. It accepts text and images and has a 262,144-token context window. It thinks before answering by default, and you can turn thinking off for direct answers (see Reasoning).

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="qwen38-27b",
    messages=[
        {"role": "user", "content": "Explain the difference between TCP and UDP in three bullet points."}
    ],
    max_tokens=2000,
)
print(response.choices[0].message.content.strip())

The answer in content can start with blank lines ("\n\n") when the model has reasoned first, so trim it before display.

Reasoning

Qwen 3.8 27B thinks by default. Read the trace from choices[].message.reasoning_content on chat, or from the reasoning output item on Responses.

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

Reasoning tokens are billed as output tokens and count against max_tokens / max_output_tokens.

Structured output

Turn thinking off when you use structured output. With thinking on, the JSON is written into the reasoning trace and content comes back null on chat (on Responses, the output has only a reasoning item).

response = client.chat.completions_v2(
    model="qwen38-27b",
    reasoning_effort="none",
    response_format={"type": "json_object"},
    messages=[
        {"role": "user", "content": "Return JSON with keys city and country for the capital of Japan."}
    ],
)

Image input

Send images as base64 data URIs in image_url content parts on chat, or input_image parts on Responses. Remote http(s) image URLs return 400. Several images can go in one message; the model reads them in order.

Supported parameters

400 is invalid_request_error.

Known limitations

• Text and image input only. Video and audio parts return 400.
• On Chat Completion V2, a developer message returns 400 invalid_request_error. Use a system message instead. Responses accepts developer input items.
• With thinking on, structured output lands in the reasoning trace instead of content. Set reasoning_effort: "none" (see Structured output).
• Reasoning and the final answer share the max_tokens / max_output_tokens budget.

Limits and errors