Qwen3-Coder-Next: The Best Self-Hostable Coding Model

Quick verdict

Qwen3-Coder-Next is the right starting point for any developer who wants serious local AI coding without sending code to a third party. It runs on a single 80 GB GPU (or 2× consumer RTX 5090), sits a few points below Anthropic's 2025 Sonnet and Opus models on the SWE-Bench scaffolds in its technical report, and costs nothing per token after the hardware investment.

Where it loses: hardest reasoning, novel algorithms, multi-file refactors with deep dependencies. For those, you keep a Claude or GPT API account on standby and route the hardest tasks there. What that saves depends on how much of your work is routine.

Specs at a glance

Hardware requirements

Sizes from the Ollama library (read November 2, 2026). We give no tokens-per-second figures here: this is a Mixture-of-Experts model, and a memory-bandwidth ceiling does not describe MoE generation speed well. Ollama has no Q5 or Q6 build of this model; the GGUF repository on Hugging Face has more quantizations if you need a size between 52 and 85 GB.

Local setup

Ollama (10-minute install)

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull Qwen3-Coder-Next at Q4_K_M (52 GB download; qwen3-coder-next:q8_0 is 85 GB)
ollama pull qwen3-coder-next

# Test it interactively
ollama run qwen3-coder-next "Write a Python function that detects palindromes"

# Or expose it as an OpenAI-compatible API
ollama serve  # listens on :11434

vLLM (production / multi-user)

pip install vllm

# Serve Qwen's FP8 weights (~80 GB) on 2× H100 80GB, 256K context, OpenAI-compatible.
# The BF16 weights (~160 GB) do not fit a single 80 GB GPU.
python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3-Coder-Next-FP8 \
  --tensor-parallel-size 2 \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.92 \
  --port 8000

# Test from any OpenAI client:
# openai.OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
# completion(model="Qwen/Qwen3-Coder-Next-FP8", messages=[...])

llama.cpp (max control)

# Build with CUDA
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build --config Release -j

# Download the official GGUF
huggingface-cli download Qwen/Qwen3-Coder-Next-GGUF \
  --include "Qwen3-Coder-Next-Q4_K_M/*" --local-dir models

# Serve (point at the first shard; llama.cpp loads the rest)
./build/bin/llama-server -m models/Qwen3-Coder-Next-Q4_K_M/Qwen3-Coder-Next-Q4_K_M-00001-of-00004.gguf \
  -ngl 99 -c 65536 --host 0.0.0.0 --port 8080

Cursor / Continue / Aider integration

Once Ollama or vLLM is serving on :11434 or :8000, you can use Qwen3-Coder-Next inside any major AI coding tool. The trick: every tool supports an “OpenAI-compatible custom endpoint”.

Base URL: http://localhost:11434/v1
API Key: dummy
Model name: qwen3-coder-next

{
  "models": [{
    "title": "Qwen3-Coder-Next",
    "provider": "openai",
    "model": "qwen3-coder-next",
    "apiBase": "http://localhost:11434/v1",
    "apiKey": "dummy"
  }]
}

export OPENAI_API_BASE=http://localhost:11434/v1
export OPENAI_API_KEY=dummy
aider --model openai/qwen3-coder-next your-file.py

Coding benchmarks

Every number below comes from the Qwen3-Coder-Next technical report, which ran all models through the same agent scaffold for each column. That makes the rows comparable with each other, which vendor-by-vendor leaderboard numbers usually are not. Scores are vendor-reported, not independent.

The report also lists non-agentic coding scores, compared only with Qwen's own models: LiveCodeBench v6 58.93 (Qwen3-Coder-480B-A35B: 44.93), EvalPlus 86.56 (480B: 86.66), MultiPL-E 88.23 (480B: 88.00) and a Codeforces rating of 2100 (480B: 1800). Aider polyglot, HumanEval and MBPP scores are not published separately for Qwen3-Coder-Next.

Source: Qwen3-Coder-Next Technical Report (arXiv 2603.00729, February 28, 2026), Tables 3-6, read November 2, 2026. Some third-party listings quote other Verified figures for the same model (different scaffolds or runs), so always check which harness a number comes from. · Qwen3-Coder-Next on Hugging Face

When to pick Qwen3-Coder-Next

• ✓You write code professionally and want a self-hosted Cursor/Aider setup.
• ✓Your code or IP cannot leave your network (regulated industries, defense, IP-sensitive).
• ✓You have or can buy a single 80 GB GPU, 2× RTX 5090, or a Mac with 64 GB+ of unified memory.
• ✓You want a hybrid setup: local for routine work, a hosted API for the hardest tasks.
• ✓You're building a coding-AI product and need an Apache-2.0-licensed model in the loop.

When to use something else

• →You want a hosted frontier model for the hardest tasks → see which Claude model to use for coding (Sonnet for value, Opus for the hardest work), or compare vendors in best AI coding models.
• →Single high-end consumer GPU with no MoE complexity → Qwen3.6-27B (dense 27B, fits a 24-32 GB card at Q4, simpler deployment).
• →You need 1M context for whole-monorepo work → DeepSeek V4 or Gemini 3.1 Pro API.
• →Mixed coding + general work, more hardware → DeepSeek V4-Flash.

Frequently asked questions