Qwen (Alibaba Cloud) is one of the most actively iterated open-weight model families in AI -- evolving from Qwen 2.5 through Qwen 3 to Qwen3-Coder-Next (February 2026) in just over a year, with sizes spanning 0.5B to 235B parameters, Apache 2.0 licensing, and the broadest multilingual support (29+ languages) of any open-weight family.
Why It's in Trial
Qwen earns Trial through breadth, velocity, and ecosystem integration:
- Rapid iteration: Qwen 2.5 (Sep 2024) -> Qwen 3 (Apr 2025) -> Qwen 3-Max (Sep 2025) -> Qwen3-Coder-Next (Feb 2026) -> Qwen 3.6-Plus (Apr 2026) -- faster release cadence than any other open-weight family
- Broadest size range: From 0.5B (edge/mobile) to 235B MoE -- no other family covers this full spectrum
- Strong coding: Qwen3-Coder-Next (80B total / 3B active MoE) reaches 70.6–71.3% SWE-bench Verified depending on scaffold; Qwen 2.5-Coder-32B scored 37.2% LiveCodeBench (beat GPT-4o's 29.2%) and 36.5% SWE-bench Verified
- Apache 2.0 licensing for most variants -- unrestricted commercial use
- Foundation for R1 distillation: DeepSeek chose Qwen 2.5 as the base for several R1-Distill variants (1.5B, 7B, 32B), validating its architecture quality
- Ecosystem presence: Referenced across this radar -- Ollama (tool calling support), Groq, Hugging Face, ACP (Qwen Code agent), Lemonade Server
- 29+ languages: Strongest multilingual coverage in the open-weight tier
The Model Family
| Model | Parameters | Release | Strength |
|---|---|---|---|
| Qwen 3.6-Plus | Proprietary API | Apr 2026 | 1M context, agentic coding focus, $0.29/M input tokens |
| Qwen3-Coder-Next | 80B total (3B active MoE) | Feb 2026 | Agentic coding: 70.6–71.3% SWE-bench Verified (Apache 2.0) |
| Qwen 3 | Dense: 0.6B-32B; Sparse: 30B (3B active), 235B (22B active) | Apr 2025 | Apache 2.0, dense + sparse |
| Qwen 3-Max | — | Sep 2025 | API flagship |
| Qwen 3-Max-Thinking | — | Jan 2026 | Reasoning variant |
| Qwen 2.5-Coder-32B | 32B | Nov 2024 | Code specialist, 37.2% LiveCodeBench |
| Qwen 2.5-Max | 236B MoE (57B active) | Jan 2025 | 90% HumanEval, $2/M tokens |
| Qwen 2.5-Omni | Multimodal | 2025 | Text, images, audio, video + speech synthesis |
Coding Capabilities
Qwen3-Coder-Next (80B total / 3B active, released February 28, 2026) is the family's frontier coding model — a sparse MoE architecture (qwen3_next) that achieves frontier-class SWE-bench Verified results while activating only 3B parameters per inference pass (arXiv 2603.00729):
| Benchmark | Qwen3-Coder-Next (SWE-Agent) | Qwen3-Coder-Next (OpenHands) |
|---|---|---|
| SWE-bench Verified | 70.6% | 71.3% |
Qwen 2.5-Coder (32B dense) remains the most widely deployed variant, trained on 5.5 trillion tokens (45% code, 55% natural language) across 92 programming languages:
| Benchmark | Qwen 2.5-Coder-32B | GPT-4o | Claude 3.5 Sonnet |
|---|---|---|---|
| LiveCodeBench | 37.2% | 29.2% | — |
| SWE-bench Verified | 36.5% | 23.6% | 33.4% |
| McEval (40+ languages) | 65.9 | — | — |
Qwen Code Agent
Qwen Code is listed in the ACP (Agent Client Protocol) registry with native support -- see the Agent Client Protocol entry. This positions Qwen alongside Kimi CLI and Mistral Vibe as one of the Chinese-origin coding agents in the ACP ecosystem.
Cautions
- Same data sovereignty considerations as other Chinese-origin models (Alibaba Cloud)
- The 3B and 72B variants use a Qwen-specific license rather than Apache 2.0 -- check licensing per variant
- Qwen3-Coder-Next's SWE-bench Verified scores (70.6–71.3%) are from the official technical report (arXiv 2603.00729); independent third-party leaderboard replication not yet confirmed
Key Characteristics
| Property | Value |
|---|---|
| Size range | 0.5B to 235B (MoE) |
| Latest generation | Qwen 3.6-Plus (April 2026, proprietary API); Qwen3-Coder-Next (February 2026, open-weights, Apache 2.0) |
| License | Apache 2.0 (most variants) |
| Languages | 29+ natural languages, 92 programming languages |
| Provider | Alibaba Cloud |
| Weights | Hugging Face: Qwen |