Qwen3.8-27B-FP8 API Documentation

Qwen3.8-27B-FP8 is a 27-billion-parameter dense vision-language model designed for coding, professional work, research, and long-horizon agentic tasks. The FP8 checkpoint uses fine-grained block quantization (block size 128) and retains performance close to the original model while reducing memory and improving inference efficiency.

Key Capabilities

Advanced reasoning and coding: Strong performance on software engineering, mathematics, research, and multi-step professional tasks.
Agentic execution: Improved planning, tool use, and handling of environment feedback for long-running workflows.
Native multimodal understanding: Accepts text, images, and videos for document analysis, visual reasoning, and other vision-language tasks.
Flexible thinking control: Thinking is enabled by default. Reasoning depth can be adjusted with reasoning_effort, and reasoning context can be retained between turns with preserve_thinking.
Long context: Supports a native context length of 262,144 tokens and can be extended up to 1,000,000 tokens when the serving configuration enables it.

Model Specifications

Item	Value
Architecture	Dense causal language model with a vision encoder
Parameters	27B
Quantization	Fine-grained FP8, block size 128
Native context length	262,144 tokens
Extended context	Up to 1,000,000 tokens, subject to serving configuration
Modalities	Text, image, and video input; text output
License	Apache 2.0

Endpoint

Send OpenAI-compatible Chat Completions requests to:

text Copy

POST https://us-01.bytecompute.ai/v1/chat/completions

Authenticate every request with your ByteCompute API key:

http Copy

Authorization: Bearer

Use the following model identifier in the request body:

text Copy

Qwen3.8-27B-FP8

Request Parameters

Parameter	Type	Required	Description
`model`	`string`	Yes	Must be `"Qwen3.8-27B-FP8"`.
`messages`	`array`	Yes	OpenAI-compatible chat messages. Message content may be text or multimodal content parts.
`max_tokens`	`integer`	No	Maximum number of tokens to generate.
`temperature`	`number`	No	Sampling randomness. Qwen recommends `1.0` in thinking mode and `0.7` in non-thinking mode.
`top_p`	`number`	No	Nucleus sampling threshold. Qwen recommends `0.95` in thinking mode and `0.80` in non-thinking mode.
`stream`	`boolean`	No	When `true`, returns Server-Sent Events (SSE) chunks.
`reasoning_effort`	`string`	No	Controls reasoning depth: `xhigh` (default), `medium`, or `low`.
`chat_template_kwargs`	`object`	No	Use `enable_thinking` and `preserve_thinking` to control reasoning behavior.

Thinking Control

Thinking mode is enabled by default. For complex coding, research, and agentic tasks:

json Copy

{
  "reasoning_effort": "xhigh",
  "chat_template_kwargs": {
    "enable_thinking": true,
    "preserve_thinking": true
  }
}

For a faster direct answer without a reasoning block:

json Copy

{
  "chat_template_kwargs": {
    "enable_thinking": false
  }
}

Recommended Sampling

Mode	Recommended settings
Thinking	`temperature=1.0`, `top_p=0.95`, `reasoning_effort=xhigh`
Non-thinking	`temperature=0.7`, `top_p=0.80`, `enable_thinking=false`

Parameter availability and maximum usable context can vary with the active inference deployment.

Qwen3.8-27B-FP8

Input

Output