Model · Alibaba · Code
Qwen3-Coder-Next
Alibaba's open-weight coding-agent model for local development — an 80B mixture-of-experts with only 3B active per token, 256K context, and built to drive agent tools such as Claude Code, Cline and Qwen Code.
- Modality
- Code
- License
- Apache 2.0 (Open)
- Parameter size
- 80B total, 3B active (MoE, 512 experts)
- Context window
- 262,144 tokens 262,144 tokens natively. Alibaba suggests cutting to 32,768 if you run out of memory.
- Released
- February 3, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- text
What you need to run it
- Mac
- Yes. Any Apple Silicon Mac (M1 or newer) with at least 64 GB (96 GB recommended) of memory. Newer chips are faster, but memory is what decides whether it runs.
- Windows / Linux PC
- NVIDIA works best. A graphics card with at least 48 GB (80 GB recommended) of video memory.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
Runs on a 64 GB or larger Mac through MLX, LM Studio or Ollama, and is quick because only 3B parameters are active per word.
Numbers are for: 4-bit build of the 80B model (about 45 GB). Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Apache 2.0 — no commercial restrictions
- Only 3B parameters active per token, so it runs fast for its size
- Trained for long-horizon agent work: tool use and recovering from failed commands
- Works inside existing coding agents — Claude Code, Qwen Code, Cline, Kilo, Trae and others
- Fill-in-the-middle supported for editor autocomplete
- FP8 and GGUF builds published by Alibaba
Weaknesses
- All 80B parameters must sit in memory: roughly a 64 GB+ Mac or a large GPU (our estimate from the size, not an Alibaba figure). On a 24 GB graphics card, use the smaller Qwen3-Coder-30B-A3B instead
- Non-thinking only — it does not produce a reasoning trace
- Needs recent SGLang (0.5.8+) or vLLM for its hybrid-attention architecture and tool parser
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | Apache 2.0; FP8 and GGUF variants also published |
| Ollama | local | Quantised builds; LM Studio, MLX-LM and llama.cpp also supported |
| OpenRouter | hosted-api | API key required |
Used in solutions
Official sources
- Model card model
- GitHub github
- Qwen blog announcement
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: size of mlx-community/Qwen3-Coder-Next-4bit.
- — Initial entry. Added 2026-10-28 from the popularity sweep: 16.8k GitHub stars (QwenLM/Qwen3-Coder), 465k Hugging Face downloads/month plus 813k for the FP8 build. Release date 2026-02-03 is from press coverage of Alibaba's announcement — the model card gives no date and the Hugging Face repo was created 2026-01-30. Apache 2.0, 262,144-token context, free to self-host.
Esc