Qwen3-Coder
Alibaba's open-weights coding model family plus the Apache 2.0 Qwen Code terminal agent, separate artifacts from the hosted Qwen Chat product. Qwen3-Coder-Next runs 80B total, 3B active.
Overview
Two things ship under this name and separating them matters. Qwen3-Coder is the open-weights coding model family from Alibaba's Qwen team. Qwen Code is the Apache 2.0 terminal agent built to drive it. Neither is Qwen Chat, which is a hosted consumer surface. These are artifacts you download and run, which is the entire point for anyone who needs to know where their source code goes.
The current flagship for local work is Qwen3-Coder-Next, released 3 February 2026. It carries 80 billion total parameters but activates only 3 billion per token through a 512-expert mixture with 10 experts live at a time, and it takes 262,144 tokens of context natively. Published results are 70.6% on SWE-bench Verified, 44.3% on SWE-bench Pro and 36.2% on Terminal-Bench 2.0. At Q4_K_XL quantization it wants roughly 40GB, which puts it inside a 64GB Mac mini M4 or a two-card workstation. llama.cpp, Ollama, vLLM and SGLang all serve it. Teams with real GPU budget still have the older Qwen3-Coder-480B-A35B-Instruct as the heavier option.
The licence is Apache 2.0 on the weights and Apache 2.0 on the CLI. No user cap, no field-of-use restriction, no acceptable-use annex that quietly excludes your industry. That matters because several Chinese open-weights releases carry community licences that read as permissive until you reach clause four. Weights land on Hugging Face and ModelScope at the same time, so mainland engineers are not waiting on a mirror. The practical warning is billing. Alibaba trimmed the free Qwen OAuth quota to 100 requests a day and then closed it on 15 April 2026, so hosted use of the CLI now needs a paid key. Alibaba Cloud also operates separate mainland and international Model Studio consoles with different catalogues and different verification requirements, and confusing the two is a common first-day mistake.
Key Features
- ✓ Qwen3-Coder-Next, released 3 February 2026: 80B total parameters with 3B activated per token, 512 experts with 10 live per token, 262,144 token native context
- ✓ Reported 70.6% on SWE-bench Verified, 44.3% on SWE-bench Pro and 36.2% on Terminal-Bench 2.0
- ✓ Apache 2.0 on both the weights and the Qwen Code CLI, with no user cap and no field-of-use annex
- ✓ Qwen Code terminal agent at 27.4k GitHub stars, speaking OpenAI, Anthropic, Gemini and Qwen protocols and pointing at Ollama or vLLM for local serving
- ✓ Runs at roughly 40GB of memory at Q4_K_XL, which fits a 64GB Mac mini M4; served by llama.cpp, Ollama, vLLM and SGLang
- ✓ Heavier sibling Qwen3-Coder-480B-A35B-Instruct still available for teams with real GPU budget, plus a qwen-code-action for GitHub Actions
Where it holds
- • Cleanest licence in Chinese open-weights coding: plain Apache 2.0 on weights and tooling, so commercial deployment needs no email to a licensing desk
- • Sparse MoE design means a 3B activation footprint at 80B quality, which is the difference between a workstation and a rented H100 cluster
- • Weights publish to Hugging Face and ModelScope together, so mainland and overseas engineers get the same artifact on day one
- • The CLI is provider-agnostic, so the same agent drives local weights, Alibaba's API or an entirely different vendor
Where it breaks
- • The free Qwen OAuth tier inside Qwen Code closed on 15 April 2026, so the terminal agent now costs either an API key or hardware
- • Qwen Code is a fork of Gemini CLI and upstream conventions still show through, with some documentation pages carrying stale Gemini-era assumptions
- • Around 40GB of memory at 4-bit puts Coder-Next above a single 24GB consumer GPU, so local use means unified-memory Macs or multi-card rigs
- • Alibaba Cloud runs separate mainland and international Model Studio consoles with different model catalogues and different sign-up requirements, and picking the wrong one wastes an afternoon
My Take
A 64GB Mac mini M4 serves this at Q4_K_XL, which is the number that changes what a small team can do. Qwen3-Coder-Next arrived on 3 February 2026 with 80 billion total parameters but only 3 billion active per token, reporting 70.6% on SWE-bench Verified. Both the weights and the CLI are Apache 2.0, the fact most English coverage skips or gets wrong. The catch came on 15 April 2026 when Alibaba closed the free Qwen OAuth tier, so the terminal agent now costs an API key or costs silicon.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0
- Starting at:
- Weights and CLI are both Apache 2.0 and free to download and self-host. The free Qwen OAuth quota inside Qwen Code was cut to 100 requests per day and then closed entirely on 15 April 2026, so hosted use now needs a paid API key through Alibaba Cloud Model Studio, a third-party provider, or a local runtime such as Ollama or vLLM at zero marginal cost.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Qwen3-Coder holds and where it breaks. The fastest check is your own workload.
Visit website ↗Qwen3-Coder compared head to head
Alternatives to Qwen3-Coder
DeepSeek
freemiumChinese AI lab producing open-source LLMs that rival proprietary models at a fraction of the cost, with strong reasoning and coding abilities.
Ling 3.0
open-sourceAnt Group's open-weight MoE family (百灵, Bailing): Ling-3.0-flash at 124B total and 5.1B active, a 7.9B tiny variant and trillion-parameter Ring reasoning models, all MIT.
MiniMax M3
freemiumMiniMax's open-weight flagship LLM (稀宇科技), distinct from the Hailuo video product: 428B mixture-of-experts, 23B active per token, 1M context and native image and video input.
CodeFuse
open-sourceAnt Group's open-source code intelligence stack (蚂蚁 CodeFuse): an Apache 2.0 GödelScript static analysis engine, MIT-licensed Ling-Coder-Lite weights and an OpenSumi-based AI IDE.