Qwen3.5-27B-FP8 API Documentation

The Qwen3.5-27B-FP8 is a state-of-the-art Sparse Mixture-of-Experts (MoE) multimodal foundation model released by Alibaba Cloud. This variant utilizes fine-grained FP8 (8-bit Floating Point) quantization, allowing the 27-billion-parameter model to achieve elite-level reasoning while significantly reducing memory footprint and increasing inference throughput compared to BF16 models.

---

Key Features

• Efficient Hybrid Architecture: Combines Gated Delta Networks with Sparse MoE, activating only a fraction of its 27B parameters per token for high-speed inference.
• Unified Vision-Language: Natively processes both text and high-resolution images within a single transformer architecture.
• 262K Native Context: Supports 262,144 tokens natively, extensible up to 1.01 million tokens via RoPE scaling (e.g., YaRN).
• Thinking Mode: Features an internal reasoning loop (Chain-of-Thought) to solve complex logic, coding, and mathematical problems.
• Hardware Optimization: FP8 support enables high-performance serving on NVIDIA Hopper (H100), Blackwell, and consumer-grade Ada Lovelace (RTX 40-series) GPUs.

---

Request Parameters

The endpoint /v1/chat/completions accepts the following parameters in a JSON-encoded body.

Parameter	Type	Required	Description
model	string	Yes	Use "Qwen/Qwen3.5-27B-FP8".
messages	array	Yes	List of message objects. Supports text and image inputs (via URL or base64).
max_tokens	integer	No	Limits generated response length. Suggested max: 8,192.
temperature	float	No	0.0 to 1.5. Use 0.6 for standard tasks; 1.0+ for creative reasoning.
stream	boolean	No	If true, tokens are delivered via Server-Sent Events (SSE).

---

Advanced Parameters

Parameter	Type	Default	Description
enable_thinking	boolean	true	Enables the <think> reasoning block. Recommended for complex tasks.
top_p	float	0.95	Nucleus sampling: limits choices to the top 95% cumulative probability.
presence_penalty	float	0.0	Range: -2.0 to 2.0. Penalizes repeated topics.
stop	array	null	Up to 4 sequences where the API will stop generating tokens.
language-model-only	boolean	false	(Serving-level) Disables vision encoder to save VRAM for KV cache.

---

Model Specifications Summary

Metric	Specification
Architecture	Sparse Mixture-of-Experts (MoE)
Quantization	FP8 (Fine-grained, Block size 128)
Context Length	262,144 Tokens (Native)
Modalities	Text, Vision (Image)
Language Support	201+ Languages