Qwen3-Coder: Alibaba's Best Coding AI
Qwen3-Coder is Alibaba's dedicated coding model family featuring the 480B flagship (35B active, rivaling Claude Sonnet) and the Coder-Next 80B (3B active, runs on consumer hardware). Both models dominate SWE-bench and agentic coding tasks while being free to run locally.
Overview: Two Models, One Mission
Alibaba's Qwen team released Qwen3-Coder in two variants optimized for different hardware tiers. The 480B flagship (July 2025) competes with proprietary models like Claude Sonnet on real-world software engineering tasks. The Coder-Next (February 2026) brings that capability to consumer hardware by activating just 3B parameters per token from an 80B MoE model with 512 experts.
Both models were trained on over 7.5 trillion tokens with 70%+ code-focused data, giving them deep understanding of programming languages, frameworks, and real-world codebases. They support agentic coding workflows where the model autonomously navigates repositories, writes code, runs tests, and iterates on solutions.
Qwen3-Coder 480B-A35B
- Parameters: 480B total, 35B active per token
- Experts: 160 total, 8 active per token
- Layers: 62 transformer layers, GQA
- Context: 262K native, 1M with YaRN
- License: Qwen License (commercial OK)
- Best for: Production deployments, maximum quality
Qwen3-Coder-Next 80B-A3B
- Parameters: 80B total, 3B active per token
- Experts: 512 total, 10 + 1 shared active
- Layers: 48 layers, hybrid DeltaNet + MoE
- Context: 256K tokens
- License: Apache 2.0 (fully open)
- Best for: Local development, consumer GPUs
Architecture Deep Dive
480B Flagship: Classic MoE at Scale
The 480B model uses a traditional Mixture-of-Experts transformer with 160 expert modules across 62 layers. Each token routes through 8 experts via top-k routing, activating 35B of the total 480B parameters. It uses grouped-query attention (GQA) for memory-efficient inference and supports native context of 262,144 tokens, extendable to 1 million tokens via YaRN positional encoding.
Coder-Next: Hybrid Architecture Innovation
Coder-Next introduces a groundbreaking hybrid architecture combining three mechanisms: Gated DeltaNet (linear attention for efficient long-range processing), Gated Attention (standard attention for precise local patterns), and MoE (512 experts with 10 active + 1 shared per token). This hybrid design achieves the coding quality of 30B+ dense models while activating only 3B parameters, enabling 10x higher throughput than comparably capable models.
| Specification | 480B Flagship | Coder-Next 80B |
|---|---|---|
| Total Parameters | 480B | 80B |
| Active per Token | 35B | 3B |
| Expert Count | 160 (8 active) | 512 (10+1 active) |
| Layers | 62 | 48 |
| Hidden Size | Unknown | 2,048 |
| Context Length | 262K (1M with YaRN) | 256K |
| Training Tokens | 7.5T+ (70% code) | 7.5T+ (70% code) |
| Architecture | MoE Transformer | Hybrid DeltaNet + MoE |
| License | Qwen License | Apache 2.0 |
| Release Date | July 2025 | February 2026 |
Benchmarks
Qwen3-Coder dominates open-source coding benchmarks. The Coder-Next variant is especially remarkable ā scoring 70.6% on SWE-bench Verified with just 3B active parameters, it competes with models 10-20x larger. All scores sourced from the Qwen3-Coder-Next Technical Report (arXiv 2603.00729) and official Qwen GitHub.
SWE-bench Verified Comparison
Coder-Next Multi-Benchmark Profile
| Benchmark | Coder-Next | 480B | DeepSeek-V3.2 | GLM-4.7 |
|---|---|---|---|---|
| SWE-bench Verified | 70.6% | 69.6% | 62.3% | 63.7% |
| SWE-bench Multilingual | 62.8% | ā | 62.3% | 63.7% |
| SWE-bench Pro | 44.3% | ā | 40.9% | 40.6% |
| Terminal-Bench 2.0 | 36.2% | ā | ā | ā |
| Aider Benchmark | 66.2% | ā | ā | ā |
Note: The Coder-Next 70.6% SWE-bench score uses SWE-Agent scaffold. With OpenHands, it reaches 71.3%. The 480B model uses different evaluation settings. Scores are not directly comparable across scaffolds. All data from official Qwen reports as of March 2026.
Quick Start with Ollama
Coder-Next (Recommended for Local Use)
Requires Ollama v0.15.5+ for hybrid architecture support. Runs on 46GB combined memory.
480B Flagship (Enterprise)
Requires ~250GB+ memory. Best run on multi-GPU servers or cloud instances.
API Access (OpenAI-Compatible)
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:11434/v1",
api_key="ollama"
)
response = client.chat.completions.create(
model="qwen3-coder-next",
messages=[
{"role": "system", "content": "You are an expert software engineer."},
{"role": "user", "content": "Write a Redis cache decorator in Python with TTL support."}
],
temperature=0.3,
max_tokens=4096
)
print(response.choices[0].message.content)Hardware Requirements
Memory needs vary dramatically between the two variants. The Coder-Next model's ultra-sparse activation (3B of 80B) keeps inference fast, but you still need enough memory to hold all expert weights. See our quantization guide for format details.
Memory Usage by Quantization
Coder-Next (Consumer Hardware)
480B Flagship (Enterprise)
Quantization & Memory Reference
| Variant | Quantization | Memory | Hardware Example |
|---|---|---|---|
| Coder-Next | Q4_K_M | ~46 GB | RTX 4090 + 24GB RAM offload |
| Coder-Next | Q3_K_M | ~38 GB | Mac Studio M2 Ultra 64GB |
| Coder-Next | Q8_0 | ~85 GB | Mac Studio M4 Ultra 128GB |
| Coder-Next | FP16 | ~160 GB | A100 80GB x2 |
| 480B | Q4_K_M | ~250 GB | A100 80GB x4 |
| 480B | FP16 | ~960 GB | H100 cluster |
Model Comparison
How Qwen3-Coder stacks up against other coding-focused models available locally.
Choose Coder-Next When:
- ā¢ You have a single consumer GPU (24GB) + system RAM
- ā¢ You need fast iteration speed for local development
- ā¢ Agentic coding with SWE-Agent, OpenHands, or Aider
- ā¢ Privacy-sensitive code that can't go to the cloud
- ā¢ Apache 2.0 license requirements
Choose 480B Flagship When:
- ā¢ You have enterprise GPU infrastructure (A100/H100)
- ā¢ Maximum quality matters more than cost
- ā¢ Complex multi-file refactoring across large codebases
- ā¢ Need 1M+ token context with YaRN extension
- ā¢ Replacing Claude Sonnet API with self-hosted alternative
Use Cases
Agentic Software Engineering
Autonomous bug fixing and feature implementation. Qwen3-Coder-Next scores 70.6% on SWE-bench Verified ā solving 7 out of 10 real GitHub issues end-to-end without human intervention.
Local AI Coding Assistant
Replace GitHub Copilot with a fully private, self-hosted coding assistant. Works with Continue.dev, Aider, and any OpenAI-compatible editor plugin via Ollama's API.
Code Review & Refactoring
Analyze entire repositories (256K+ context) for bugs, security vulnerabilities, and code smells. The MoE architecture enables fast scanning across large codebases.
Multi-Language Development
Trained on 7.5T+ tokens across all major programming languages. Strong performance on Python, TypeScript, Rust, Go, Java, C++, and emerging languages.
Terminal Automation
Scores 36.2% on Terminal-Bench 2.0 ā understanding shell commands, DevOps workflows, and system administration tasks for automated infrastructure management.
Enterprise Code Migration
Use the 480B flagship with 1M token context to analyze and migrate legacy codebases. Handles framework upgrades, API migrations, and language conversions.
Advanced Setup
Custom Modelfile for Coding Tasks
# Save as Modelfile.qwen3-coder FROM qwen3-coder-next PARAMETER temperature 0.2 PARAMETER top_p 0.9 PARAMETER num_ctx 131072 PARAMETER repeat_penalty 1.05 PARAMETER stop "<|endoftext|>" PARAMETER stop "<|im_end|>" SYSTEM """You are an expert software engineer. When given a task: 1. Analyze the existing code structure 2. Plan your changes before implementing 3. Write clean, tested, production-ready code 4. Explain key decisions briefly Always follow existing code conventions and patterns."""
Multi-GPU Setup for 480B
# vLLM multi-GPU deployment for 480B
pip install vllm>=0.6.0
python -m vllm.entrypoints.openai.api_server \
--model Qwen/Qwen3-Coder-480B-A35B-Instruct \
--tensor-parallel-size 4 \
--max-model-len 262144 \
--gpu-memory-utilization 0.95 \
--trust-remote-code \
--port 8000Using with Aider (Coding Agent)
# Install Aider
pip install aider-chat
# Point Aider at your local Ollama instance
export OLLAMA_API_BASE=http://localhost:11434
# Start coding session with Qwen3-Coder-Next
aider --model ollama/qwen3-coder-next
# Or use the 480B via API
aider --model openai/qwen3-coder:480b \
--openai-api-base http://localhost:11434/v1 \
--openai-api-key ollamaSources
- Qwen3-Coder-Next Technical Report (arXiv 2603.00729, March 2026) ā Architecture, benchmarks, and training details
- QwenLM/Qwen3-Coder GitHub Repository ā Official code, model cards, and documentation
- Qwen Blog: Qwen3-Coder-Next Announcement ā Official release blog post and capabilities
- Qwen3-Coder-480B-A35B-Instruct on Hugging Face ā Model weights and technical specifications
- Qwen3-Coder on Ollama Library ā Ollama model variants, sizes, and download
Frequently Asked Questions
What is Qwen3-Coder and who made it?
Qwen3-Coder is Alibaba's Qwen team's dedicated coding model family, released in mid-2025. The flagship 480B model (35B active MoE) rivals Claude Sonnet on agentic coding tasks. The smaller Qwen3-Coder-Next (80B total, 3B active) was released February 2026 for local development, running on just 46GB of memory while matching models 10-20x larger.
Can I run Qwen3-Coder on my computer?
Yes. Qwen3-Coder-Next (80B, 3B active) runs on ~46GB combined memory ā a 24GB GPU (RTX 4090) plus system RAM handles it well. At Q4_K_M quantization via Ollama, it fits comfortably on consumer hardware. The full 480B model needs ~250GB+ memory, requiring enterprise GPUs or multi-GPU setups.
How does Qwen3-Coder compare to DeepSeek Coder V2?
Qwen3-Coder 480B significantly outperforms DeepSeek Coder V2 (236B) on SWE-bench Verified (69.6% vs ~49%), LiveCodeBench, and agentic coding tasks. Qwen3-Coder-Next (3B active) also surpasses DeepSeek Coder V2 Lite while using far fewer active parameters. The Qwen3 series represents a generational leap in open-source coding models.
What is the Qwen3-Coder-Next variant?
Qwen3-Coder-Next is a highly efficient 80B MoE model that activates only 3B parameters per token using 512 experts (10 active + 1 shared per token). It uses a hybrid architecture combining Gated DeltaNet, Gated Attention, and MoE across 48 layers. Despite its small active parameter count, it scores 70.6% on SWE-bench Verified ā comparable to the much larger 480B flagship.
What context length does Qwen3-Coder support?
The flagship Qwen3-Coder 480B supports 262,144 tokens natively, extendable to 1M tokens using YaRN scaling. Qwen3-Coder-Next supports 256K tokens. Both models handle full repository-scale codebases in a single context window, making them excellent for agentic coding workflows that need to understand entire projects.
How do I run Qwen3-Coder with Ollama?
For the Next variant: ollama pull qwen3-coder-next followed by ollama run qwen3-coder-next. For the 480B: ollama pull qwen3-coder:480b. Note that Qwen3-Coder-Next requires Ollama v0.15.5 or newer to support its hybrid MoE/SSM architecture. Both models are available in multiple quantization levels.
Is Qwen3-Coder open source?
Qwen3-Coder 480B is released under the Qwen License (permissive, allows commercial use). Qwen3-Coder-Next uses the Apache 2.0 license, which is fully open source with no restrictions. Both models have open weights available on Hugging Face and can be freely downloaded, modified, and deployed.
What are Qwen3-Coder's best benchmarks?
Qwen3-Coder-Next scores: SWE-bench Verified 70.6% (71.3% with OpenHands), SWE-bench Multilingual 62.8%, SWE-bench Pro 44.3%, Terminal-Bench 2.0 36.2%, and Aider 66.2%. The 480B flagship scores 69.6% on SWE-bench Verified and leads open models on agentic coding and browser-use tasks, rivaling Claude Sonnet.
Was this helpful?
Related Guides
Continue your local AI journey with these comprehensive guides
Written by Pattanaik Ramswarup
AI Engineer & Dataset Architect | Creator of the 77,000 Training Dataset
I've personally trained over 50 AI models from scratch and spent 2,000+ hours optimizing local AI deployments. My 77K dataset project revolutionized how businesses approach AI training. Every guide on this site is based on real hands-on experience, not theory. I test everything on my own hardware before writing about it.