Qwen3 Coder 480B A35B
Active Parameters
480B
Context Length
262K
Modality
Text
Architecture
Mixture of Experts (MoE)
License
Apache 2.0
Release Date
22 Jul 2025
Knowledge Cutoff
Dec 2024
API Pricing (per 1M)
Input: $1.50 · Output: $7.50
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
58x RTX 4090
24GB VRAM
Datacenter
15x NVIDIA A100
80GB VRAM
Apple Silicon
12x Apple M3 Max
128GB VRAM
262,144 tokens
Consumer
63x RTX 4090
24GB VRAM
Datacenter
17x NVIDIA A100
80GB VRAM
Apple Silicon
13x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#124
| Benchmark | Score | Rank |
|---|---|---|
Software Engineering | 0.70 | 16 |
Web Development | 1274 | 90 |
General Text | 1387 | 104 |
Intelligence Index | 0.12 | 203 |
Rankings
Overall Rank
#124
Coding Rank
#86
About Qwen3 Coder 480B A35B
Qwen3 Coder 480B A35B is Alibaba's specialized agentic coding MoE model activating 35B parameters for repository-wide software development. Trained on 7.5T tokens with Agent RL, it natively powers terminal agents and automated PR generation over a 1M context.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
96
Key-Value Heads
8
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
10,000,000
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
6,144
Number of Layers
62
FFN Intermediate Size (Dense)
2,560
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
151,936
Mixture of Experts
Total Expert Parameters
35.0B
Number of Experts
160
Active Experts
8
Shared Experts
-
FFN Intermediate Size (per Expert)
2,560
Dense Layers Before MoE
-
About Qwen 3
The Alibaba Qwen 3 model family comprises dense and Mixture-of-Experts (MoE) architectures, with parameter counts from 0.6B to 235B. Key innovations include a hybrid reasoning system, offering 'thinking' and 'non-thinking' modes for adaptive processing, and support for extensive context windows, enhancing efficiency and scalability.