Qwen3-Coder 30B and 480B: Run Alibaba's Coding Models Locally

Overview: Two Sizes, One Model Family

Architecture Deep Dive

480B Flagship: Classic MoE at Scale

30B-A3B: The Same Design, Scaled Down

Benchmarks

Qwen published the 480B's headline SWE-bench Verified score, 69.6% with the OpenHands scaffold at 500 turns, in the chart of its July 2025 announcement. The most detailed like-for-like numbers come later, from the Qwen3-Coder-Next Technical Report (arXiv 2603.00729), which also lists the 480B on non-agentic coding tests. Qwen published the 30B-A3B's own scores only as a chart image on its model card, but the official swebench.com board lists Qwen3-Coder 30B at 51.6% on SWE-bench Verified, run inside the OpenHands agent (read October 15, 2026).

SWE-bench Verified with OpenHands

480B: Qwen3-Coder announcement chart (500 turns). Other rows: Qwen3-Coder-Next technical report, Table 3. The two runs are from different reports, so treat a gap of a point or two as noise.

480B Non-Agentic Coding Profile

Qwen3-Coder-480B-A35B in the Qwen3-Coder-Next technical report, Table 6: EvalPlus 86.66, MultiPL-E 88.00, CRUXEval 92.13, LiveCodeBench v6 44.93, OJBench 14.98, Codeforces rating 1800.

Quick Start with Ollama

30B-A3B (Recommended for Local Use)

Install Ollama, then pull the 30B. The plain qwen3-coder tag points at the same 19GB Q4_K_M build.

Other 30B builds: qwen3-coder:30b-a3b-q8_0 (32GB) and qwen3-coder:30b-a3b-fp16 (61GB). The 80B successor is qwen3-coder-next (52GB); see the Qwen3-Coder-Next guide.

480B-A35B Flagship (Server)

The Q4_K_M build is 290GB, so it needs roughly 300GB+ of GPU or unified memory before any context.

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:30b",
    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

Both models are Mixture-of-Experts: only a small part runs per token, which keeps generation fast, but every expert has to be loaded, so memory is set by the total size. The sizes below are the Ollama downloads; add the KV cache for your context on top. See our quantization guide for what Q4_K_M and Q8_0 mean. We give no tokens-per-second figures: a memory-bandwidth ceiling does not describe MoE generation speed well.

Ollama Download Size by Quantization (GB)

Model Comparison

How Qwen3-Coder compares with other open-weight coding models you can pull from Ollama, by size and the hardware each needs. For benchmark scores, see the table above and our best Ollama model for coding guide.

Between the two sits Qwen3-Coder-Next: 52GB at Q4, so it needs 64GB of memory, with a published 70.6% SWE-bench Verified (SWE-Agent).

Use Cases

Advanced Setup

Custom Modelfile for Coding Tasks

# Save as Modelfile.qwen3-coder
FROM qwen3-coder:30b

PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20
PARAMETER repeat_penalty 1.05
# 32K context fits a 24GB GPU next to the 19GB weights (~3GB KV cache)
PARAMETER num_ctx 32768

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."""

Sampling values follow the recommended settings on the Qwen3-Coder model cards (temperature 0.7, top_p 0.8, top_k 20, repetition penalty 1.05).

Multi-GPU Setup for 480B

# vLLM on 8× H100 80GB with the official FP8 weights (~480GB)
pip install -U vllm

vllm serve Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 \
    --tensor-parallel-size 8 \
    --max-model-len 131072 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml \
    --port 8000

Using 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 a coding session with the 30B
aider --model ollama_chat/qwen3-coder:30b

# Or use the 480B served by vLLM (OpenAI-compatible)
aider --model openai/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 \
    --openai-api-base http://localhost:8000/v1 \
    --openai-api-key dummy

Sources

Frequently Asked Questions