Qwen3

2025-04-29 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba Qwen releases Qwen3, eight open-weights models under Apache 2.0 (two MoE, 235B-A22B and 30B-A3B, and six dense from 0.6B to 32B) that combine a thinking mode and a direct mode in a single model.

### Context Qwen2.5 (September 2024) offered dense models and a separate reasoning model, QwQ. In spring 2025, reasoning models (o1, DeepSeek-R1, Gemini 2.5 Pro) required two distinct families of models depending on the use. Qwen3 aims to unify them and to open the entire lineup, from the 0.6-billion-parameter model to the 235-billion MoE. ### What the model brings - Hybrid modes: a thinking mode (step-by-step reasoning) and a non-thinking mode (fast answer), switchable with `enable_thinking` or with the /think and /no_think tags, with an adjustable thinking budget. - Pre-training on about 36 trillion tokens in 119 languages and dialects, compared with 18 trillion for Qwen2.5; PDFs are extracted with Qwen2.5-VL and synthetic math and code data are produced with Qwen2.5-Math and Qwen2.5-Coder. - Three pre-training stages: more than 30 trillion tokens at 4,000 tokens of context, 5 trillion tokens richer in STEM and code, then context extension to 32,000 tokens. - Four-stage post-training: cold start on long chains of reasoning, reinforcement on reasoning, mode fusion, general reinforcement on more than 20 tasks. - Efficiency: the MoE base models match dense Qwen2.5 models with 10% of the active parameters, and Qwen3-4B is announced at the level of Qwen2.5-72B-Instruct. ### Architecture and specifications | Model | Type | Context | |---|---|---| | Qwen3-235B-A22B | MoE, 235 billion total, 22 billion active, 128 experts (8 active), 94 layers | 128 K | | Qwen3-30B-A3B | MoE, 30 billion total, 3 billion active, 128 experts (8 active), 48 layers | 128 K | | Qwen3-32B, 14B, 8B | Dense | 128 K | | Qwen3-4B, 1.7B, 0.6B | Dense | 32 K | Apache 2.0 license (commercial use allowed, no threshold). Weights on Hugging Face, ModelScope and Kaggle; technical report arXiv 2505.09388. Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Weights | Free, Apache 2.0 | | Alibaba Cloud API qwen3-235b-a22b (current pricing consulted, not the launch pricing) | $0.7 input / $2.8 output (non-thinking), $8.4 output (thinking) per million tokens | | Required hardware | Not disclosed by the blog | ### Benchmarks Qwen3-235B-A22B in thinking mode, table from the technical report (vendor measurements). | Benchmark | OpenAI-o1 | Grok-3 Beta | Gemini 2.5 Pro | Qwen3-235B-A22B | |---|---|---|---|---| | AIME'24 | 74.3 | 83.9 | 92.0 | 85.7 | | AIME'25 | 79.2 | 77.3 | 86.7 | 81.5 | | LiveCodeBench v5 | 63.9 | 70.6 | 70.4 | 70.7 | | CodeForces (Elo) | 1891 | not published | 2001 | 2056 | | GPQA-Diamond | 78.0 | 80.2 | 84.0 | 71.1 | | BFCL v3 | 67.8 | not published | 62.9 | 70.8 | | Arena-Hard | 92.1 | not published | 96.4 | 95.6 | Source: Qwen3 technical report (arXiv 2505.09388), vendor measurements. In non-thinking mode, the same report gives Arena-Hard at 96.1 for Qwen3-235B-A22B against 85.3 for GPT-4o-2024-11-20. ### Versus closed models Against Gemini 2.5 Pro, Qwen3-235B-A22B is behind on AIME'24 (85.7 versus 92.0), GPQA-Diamond (71.1 versus 84.0) and Arena-Hard (95.6 versus 96.4), but ahead on LiveCodeBench v5 (70.7 versus 70.4), CodeForces (2056 versus 2001) and BFCL v3 (70.8 versus 62.9). Against OpenAI-o1, it is ahead on AIME'24, AIME'25, code and tool calls, and behind on GPQA-Diamond (71.1 versus 78.0). In direct mode, it surpasses GPT-4o-2024-11-20 on the four math and code tests in the report's table (for example LiveCodeBench v5, 35.3 versus 32.7). These figures are the vendor's. The main practical contribution is the Apache 2.0 license across the whole lineup, which allows royalty-free self-hosting. ### Key takeaways Qwen3 is the milestone that unifies thinking and direct answering in an open model and extends openness to the whole range of sizes. Choosing the thinking mode per task makes it possible to trade off cost and quality without changing model.

Tags : LLM, Open Source, Alibaba, Qwen, MoE, Apache 2.0, Raisonnement