Qwen3.5-122B-A10B-FP8 API Documentation

Qwen3.5-122B-A10B-FP8 is the flagship Sparse Mixture-of-Experts (MoE) multimodal foundation model from the Qwen3.5 series. By integrating 122 billion total parameters with 10 billion active parameters per token, it delivers elite-level reasoning capacity with the inference efficiency of a much smaller model.

This version is optimized with fine-grained FP8 quantization, specifically engineered for NVIDIA Blackwell and Hopper (H100/L40S) architectures to maximize throughput in complex multimodal reasoning and long-context applications.

---

Key Capabilities

• Unified Vision-Language: Natively processes both text and high-resolution images within a single transformer foundation.
• Native Thinking Mode: Features an internal reasoning loop (Chain-of-Thought) for superior performance in math, coding, and complex logic.
• Massive 262K Context: Natively supports up to 262,144 tokens, extensible to 1.01 million tokens for massive document or codebase analysis.
• Global Linguistic Coverage: Advanced optimization for 201+ languages and dialects, including industry-leading CJK and English support.
• FP8 Efficiency: Significant VRAM reduction and hardware-accelerated throughput on modern data center GPUs.

---

Request Parameters (/v1/chat/completions)

Parameter	Type	Required	Description
model	string	Yes	Must be "Qwen3.5-122B-A10B-FP8".
messages	array	Yes	Array of message objects. Supports text and image inputs.
max_tokens	integer	No	Maximum tokens to generate. Suggested range: 1 to 8,192.
enable_thinking	boolean	No	Default: true. Enables the <think> reasoning block.
temperature	float	No	Controls randomness. 0.2 for logic, 0.7 for general chat.
top_p	float	No	Nucleus sampling threshold. Default: 0.95.
stream	boolean	No	Whether to stream tokens in real-time via SSE.

---

Optimization Scenarios

Scenario	Recommended Params	Purpose
Deep Reasoning	enable_thinking: true, temp: 0.2	Best for complex math, logic, and multi-step coding problems.
Multimodal Analysis	messages: [with image], temp: 0.4	Ideal for document OCR, chart reasoning, and visual QA.
Long Doc Synthesis	max_tokens: 4096+, top_p: 0.9	Leverages the 262K context for accurate, long-form extraction.
Global Translation	temperature: 0.3	Utilizes the 201-language support for nuanced, cultural translation.