# Qwen3

> **TL;DR.** Qwen3 is Alibaba's third-generation open-weight foundation model family, launched [April 28, 2025](https://qwenlm.

- **Category:** AI / Foundation Models / Open Source
- **Stage:** established
- **Age:** 401 days
- **Origin date:** 2025-04-28
- **First detected:** 2026-04-17
- **Canonical URL:** https://earlyterms.com/term/qwen3
- **Sources:** 8 primary URLs

## Definition

Qwen3 is Alibaba's third-generation open-weight foundation model family, launched [April 28, 2025](https://qwenlm.github.io/blog/qwen3/) under Apache 2.0. It introduced a hybrid 'Thinking' / 'Non-Thinking' mode — a single model that can reason step-by-step or answer quickly — and shipped six dense sizes (0.6B-32B) plus two MoE variants. Flagship Qwen3-235B-A22B reached parity with DeepSeek-R1, o1, and Gemini-2.5-Pro on reasoning benchmarks.

A year of aggressive follow-ups built out the franchise: Qwen3-Coder (July 2025), Qwen3-Next + Qwen3-Omni (Sept 2025), Qwen3-TTS (Jan 2026), Qwen3-Max-Thinking (Jan 2026), Qwen3.5 (March 2026), Qwen3.6-Plus (April 2 2026), and the [Qwen3.6-35B-A3B MoE](https://qwen.ai/blog?id=qwen3.6-35b-a3b) on April 16, 2026 — which posted a 1,262-point HN thread and 73.4 on SWE-Bench Verified running locally.

## Example

Qwen3.6-35B-A3B activates only 3B of its 35B parameters per token, yet beats Gemma4-31B (52.0 → 73.4) on SWE-Bench Verified and posts 81.7 on MMMU — above Claude Sonnet 4.5 (79.6). Simon Willison's April 16, 2026 [pelican-on-a-bicycle test](https://simonwillison.net/2026/Apr/16/qwen-beats-opus/) concluded: 'I'm giving this one to Qwen 3.6. Opus managed to mess up the bicycle frame!' Running on a laptop, not a data center.

## Analogy

If Llama was the open Chevy of LLMs, Qwen3 is the open Toyota — quietly shipping more variants, more often, with benchmarks that close the gap to the flagships.

## Why it's emerging now

Qwen3.6-35B-A3B (April 16, 2026) crossed a threshold — laptop-runnable, 73.4 on SWE-Bench Verified, beats Claude Sonnet 4.5 on MMMU — and posted a 1,262-point HN thread in a day. Combined with an April 2 release of Qwen3.6-Plus targeted at 'real world agents', Qwen3 is the open-weight family every local-LLM and agent builder is defaulting to this month.

## Related terms

- *parent:* Qwen
- *parent:* Alibaba Cloud
- *child:* Qwen3-Coder
- *child:* Qwen3-Omni
- *child:* Qwen3-TTS
- *child:* Qwen3-Max
- *child:* Qwen3-Next
- *competitor:* Llama
- *competitor:* DeepSeek
- *competitor:* Mistral
- *competitor:* claude-opus-4-7
- *related:* lm-studio
- *alias:* 通义千问

## Sources

1. [Qwen — official launch blog](https://qwenlm.github.io/blog/qwen3/)
2. [Qwen3.6-35B-A3B announcement](https://qwen.ai/blog?id=qwen3.6-35b-a3b)
3. [Qwen3 Technical Report (arXiv)](https://arxiv.org/abs/2505.09388)
4. [Qwen3 on GitHub (QwenLM/Qwen3)](https://github.com/QwenLM/Qwen3)
5. [Simon Willison — Qwen3.6 beats Opus 4.7](https://simonwillison.net/2026/Apr/16/qwen-beats-opus/)
6. [Hacker News — Qwen3.6-35B-A3B thread (1,262 pts)](https://news.ycombinator.com/item?id=47792764)
7. [Alibaba Cloud — Qwen3.6 agentic coding blog](https://www.alibabacloud.com/blog/qwen3-6-35b-a3b-agentic-coding-power-now-open-to-all_603043)
8. [Unsloth — run Qwen 3.5 locally](https://unsloth.ai/docs/models/qwen3.5)

---
_Generated by EarlyTerms · https://earlyterms.com/term/qwen3_