Qwen3.5-27B-FP8 Documentation

The Qwen3.5-27B-FP8 is a dense, multimodal foundation model released by Alibaba Cloud in February 2026. This specific variant utilizes FP8 (8-bit Floating Point) quantization, allowing the 27-billion-parameter model to achieve high-density reasoning while maintaining low latency and a smaller memory footprint compared to its BF16 counterpart.

Key Features

• Dense Power: Unlike MoE siblings, it activates all 27B parameters for every token, yielding industry-leading scores on SWE-bench (72.4%).
• Hybrid Attention: Uses a Gated DeltaNet (linear + full attention) architecture for efficient long-context processing.
• 256K Context Window: Native support for 262,144 tokens, extensible up to 1M tokens.
• Native Multimodal: Seamlessly processes text and high-resolution images within the same context.
• Quantization Mastery: FP8 precision offers ~1.5x throughput increase on NVIDIA Blackwell and Hopper architectures.

---

Request Parameters

The following parameters must be included in the body of your POST request to the /v1/chat/completions endpoint.

Parameter	Type	Required	Description
model	string	Yes	Use "Qwen3.5-27B-FP8".
messages	array	Yes	A list of message objects (role: system/user/assistant, content: string/array).
max_tokens	integer	No	Limits the length of generated response. Default: 4096. Max: 262,144.
temperature	float	No	Controls randomness. 0.6 is recommended for standard tasks; 1.0 for reasoning.
stream	boolean	No	If true, tokens are sent as Server-Sent Events (SSE) as they are generated.

---

Optional Parameters

These parameters allow for fine-grained control over the model's behavior and output quality.

Parameter	Type	Default	Description
top_p	float	0.95	Nucleus sampling. Limits the next token choice to the top 95% cumulative probability.
top_k	integer	40	Limits choice to the Top-K most likely tokens. Reduces rare word "loops."
presence_penalty	float	0.0	Range: -2.0 to 2.0. Increases likelihood of talking about new topics.
stop	string/array	null	Up to 4 sequences where the API will stop generating further tokens.
enable_thinking	boolean	false	(Specific to reasoning providers) Enables the <think> reasoning block.

---

Parameters Summary

Metric	Specification
Supported Modalities	Text, Image
Default Context	262,144 Tokens
Output Type	Text, Markdown, Structured JSON
Language Support	201+ Languages