Alibaba · 480B (35B active) · Mixture of Experts
Largest open coding MoE — 35B active
72.3K downloads
1.3K likes
2025-07 256K context
Use Cases
code
Mixture of Experts
Total experts: 128
Active experts: 8
Active params: 35.0B
| Quant | Bits | VRAM | Quality | Status |
|---|---|---|---|---|
| Q2_K | 2 | 154.2 GB | low | — |
| Q3_K_M | 3 | 215.6 GB | moderate | — |
| Q4_K_M | 4 | 246.4 GB | good | — |
| Q5_K_M | 5 | 307.8 GB | good | — |
| Q6_K | 6 | 369.3 GB | excellent | — |
| Q8_0 | 8 | 492.2 GB | excellent | — |
| F16 | 16 | 984 GB | lossless | — |
About this model
Qwen3-Coder is the most agentic code model to date in the Qwen series.
Get started
480B
Cloud
ollama run qwen3-coder:480b-cloud
Local
ollama run qwen3-coder:480b
Running locally requires a minimum of 250GB of memory or unified memory.
30B
ollama run qwen3-coder:30b
Overview
qwen3-coder:30b offers 30B total parameters with only 3.3B activated, delivering strong performance while maintaining efficiency.
- Exceptional agentic capabilities for real-world software engineering tasks through advanced long-horizon reinforcement learning on SWE-Bench and similar benchmarks.
- Long context support with 256K tokens natively and up to 1M tokens using extrapolation methods, optimized for repository-scale understanding.
- Scaled pretraining on 7.5T tokens (70% code ratio) while preserving strong general and mathematical abilities.
- Execution-driven reinforcement learning that significantly boosts code execution success rates across diverse real-world coding tasks.