Qwen
Quick Facts
- Vendor
- Alibaba Cloud (Hangzhou)
- Released
- Qwen (2023); Qwen3 (2025)
- Current line
- Qwen3 (0.6B–235B) · Qwen3-Coder · Qwen-Agent
- License
- Apache 2.0 (most variants)
- Hosting
- Self-hosted (vLLM, Ollama); hosted via DashScope, Together, Fireworks
- Context window
- 32K–1M tokens depending on variant
- Modalities
- Text, vision (Qwen-VL), audio (Qwen-Audio)
- Architecture
- Dense and MoE variants
Summary
Qwen is Alibaba's open-weights LLM family, first released in 2023. Qwen3 (2025) shipped a tiered lineup from 0.6B to 235B parameters, with mixture-of-experts at the top and dense models filling out the smaller tiers. Qwen3-Coder is the specialized coding variant and, on most benchmarks, the strongest open model for agentic code tasks.
The Qwen-Agent framework is as important as the weights. It wraps tool calling, MCP server integration, a Python code interpreter, a browsing agent, and retrieval into a cohesive agent primitive — the same design patterns Claude Agent SDK and Stripe's Minions use, but on open infrastructure. For teams building managed AI services that can't rely on closed APIs, Qwen + Qwen-Agent is the closest thing to a turnkey stack.
Model Lineup
- Qwen3-235B-A22B — flagship MoE. Frontier-competitive general model.
- Qwen3-Coder — code-specialized variant. The reference open model for coding agents; Aider and SWE-bench scores rival closed frontier.
- Qwen3 7B / 14B / 32B — dense tiers. Sweet spot for self-hosted edge deployments.
- Qwen-Agent — agent framework. Tool calling, MCP, browser, code interpreter, RAG.
- Qwen-VL / Qwen-Audio — multimodal variants.
Where Qwen Fits
Qwen is the default open model for coding agents and for workloads where tool-use reliability matters more than raw benchmark scores. Apache 2.0 licensing sidesteps the enterprise-scale restrictions in Llama's license. Within the Claw ecosystem, Qwen3-Coder powers OpenClaw and NanoClaw — the agent runtimes that handle interior code loops on customer hardware. Qwen-Agent is the framework the ecosystem is named after.
Tradeoffs
- Provenance. Some regulated and government customers have restrictions on Chinese-origin models. Check procurement policy.
- Documentation. English documentation is solid but lags the Chinese docs. Some framework features are easier to discover by reading the source.
- Tool-use reliability is the best in the open-weights tier but still below Claude at frontier scale. Plan for retry logic on long tool chains.
Deployment Notes
Qwen3-Coder is the engine behind OpenClaw and NanoClaw. It runs on customer Mac Minis via vLLM or llama.cpp, with Qwen-Agent wrapping the agent loop and MCP bridges connecting to Claude Code Slack, FrawdBot, and the observability stack. For teams that want a pure-open stack — no closed API calls — Qwen is typically the highest-capability choice. See The Agent Infrastructure Stack for the full topology.