Skip to main content

​
About the Provider

Alibaba Cloud is the cloud computing arm of Alibaba Group and the creator of the Qwen model family. Through its open-source initiative, Alibaba has released state-of-the-art language and multimodal models under permissive licenses, enabling developers and enterprises to build powerful AI applications across diverse domains and languages.

​
Model Quickstart

This section helps you quickly get started with the Qwen/Qwen3.5-35B-A3B model on the Qubrid AI inferencing platform. To use this model, you need:
  • A valid Qubrid API key
  • Access to the Qubrid inference API
  • Basic knowledge of making API requests in your preferred language
Once authenticated with your API key, you can send inference requests to the Qwen/Qwen3.5-35B-A3B model and receive responses based on your input prompts. Below are example placeholders showing how the model can be accessed using different programming environments.
You can choose the one that best fits your workflow. 
from openai import OpenAI

# Initialize the OpenAI client with Qubrid base URL
client = OpenAI(
    base_url="https://platform.qubrid.com/v1",
    api_key="QUBRID_API_KEY",
)

# Create a streaming chat completion
stream = client.chat.completions.create(
    model="Qwen/Qwen3.5-35B-A3B",
    messages=[
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "What is in this image? Describe the main elements."
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
            }
          }
        ]
      }
    ],
    max_tokens=8192,
    temperature=0.6,
    top_p=0.95,
    stream=True
)

# If stream = False comment this out
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)
print("\n")

# If stream = True comment this out
print(stream.choices[0].message.content)

​
Model Overview

Qwen3.5-35B-A3B is the breakout model of the Qwen3.5 Medium Series and the biggest efficiency breakthrough in recent open-source AI.
  • With 35B total parameters and only 3B active per token (8.6% of total) across a 60-layer hybrid architecture, it delivers frontier-level knowledge and visual reasoning at a fraction of the compute cost.
  • It outperforms the previous generation’s Qwen3-235B-A22B on most benchmarks, as well as GPT-5 mini and Claude Sonnet 4.5 on knowledge (MMMLU) and visual reasoning (MMMU-Pro).
  • Runs on an 8GB GPU (4-bit quantization) or a 22GB Mac M-series, and supports text, image, and video input natively via early fusion.

​
Model at a Glance

FeatureDetails
Model IDQwen/Qwen3.5-35B-A3B
ProviderAlibaba Cloud (Qwen Team)
ArchitectureHybrid Gated DeltaNet + Sparse MoE Transformer — 60 layers, 3:1 linear-to-full attention ratio, 256 experts (8 routed + 1 shared per token), early fusion multimodal vision encoder, MTP speculative decoding
Model Size35B Total / 3B Active
Context Length256K Tokens (up to 1M)
Release DateFebruary 24, 2026
LicenseApache 2.0
Training DataTrillions of multimodal tokens (text, image, video) across 201 languages; RL post-training scaled across million-agent environments

​
When to use?

You should consider using Qwen3.5-35B-A3B if:
  • You need a powerful vision-language model deployable on consumer or edge hardware
  • Your application requires cost-efficient enterprise inference at scale
  • You are building agentic coding and tool-calling workflows on a budget
  • You need long-context document analysis with 256K native context
  • Your use case involves multimodal chat across text, image, and video
  • You need complex reasoning with optional thinking mode
  • You want a fully open-source model (Apache 2.0) with near-lossless 4-bit quantization

​
Inference Parameters

Parameter NameTypeDefaultDescription
StreamingbooleantrueEnable streaming responses for real-time output.
Temperaturenumber0.6Use 0.6 for non-thinking mode, 1.0 for thinking/reasoning mode.
Max Tokensnumber8192Maximum number of tokens to generate.
Top Pnumber0.95Nucleus sampling parameter.
Top Knumber20Limits token sampling to top-k candidates.
Enable ThinkingbooleanfalseToggle chain-of-thought reasoning. Set temperature=1.0 when enabled. Increases latency.

​
Key Features

  • Historic Efficiency: Beats Qwen3-235B-A22B with only 3B active parameters — 8.6% of total params per token.
  • Outperforms GPT-5 mini & Claude Sonnet 4.5: Higher scores on MMMLU (knowledge) and MMMU-Pro (visual reasoning).
  • 8GB GPU Deployment: Runs on consumer hardware via near-lossless 4-bit quantization; also supports 22GB Mac M-series.
  • 256K Native Context: Supports up to 1M tokens with extended configuration.
  • Native Multimodal: Text, image, and video via early fusion — no separate vision encoder.
  • MTP Speculative Decoding: Enhanced throughput via Multi-Token Prediction.
  • Apache 2.0 License: Fully open source with full commercial freedom.

​
Summary

Qwen3.5-35B-A3B is the efficiency breakthrough of the Qwen3.5 Medium Series, delivering frontier-level performance at a fraction of the compute cost.
  • It uses a 60-layer hybrid Gated DeltaNet + Sparse MoE architecture with 35B total and only 3B active parameters per token.
  • It outperforms Qwen3-235B-A22B, GPT-5 mini, and Claude Sonnet 4.5 on key knowledge and visual reasoning benchmarks.
  • The model supports 256K native context, optional thinking mode, and 201 languages.
  • Licensed under Apache 2.0 for full commercial use, deployable on an 8GB GPU.