Qwen3
The fifth-generation Qwen large model series from Alibaba, described as the first open-source hybrid reasoning model family. It comprises six model sizes and supports both Dense and MoE architectures; the flagship 235B-A22B model is positioned to compete with Gemini 2.5 Pro.
Core Facts
The Qwen2.5:7B model is approximately 4 GB in size after Q4 quantization.
Qwen（通义千问）是阿里巴巴达摩院开源的系列大模型，覆盖从0.5B到72B的多种参数规模，支持多语言和多模态任务
Qwen3 introduces a dynamic switching mechanism between thinking mode and non-thinking mode.
Timeline (last 90 days)
The repository is based on the paper 'The Pain Axis: LLMs Represent Self-Directed Harm and Act on It' and uses small open-source models (Qwen3, Llama 3.2, Phi-4-mini) for experiments.
On models of 4B and below, the normal-style prompt scored highest, with the plain and florid styles both dropping about 8 points, but all confidence intervals crossed zero, so the differences cannot be strictly proven significant.
On math tasks, the accuracy was 79% when the model wrote out reasoning steps, and only 11% when it gave a bare number.
In a test covering 8 local models and more than 5000 calls, the wording style of the prompt (plain/normal/florid) had almost no effect on accuracy on non-math tasks, with differences of no more than 4 percentage points for models of 8B and above.
The test covered 8 local models: Qwen3's 0.6B, 1.7B, 4B, 8B, and 14B, as well as Gemma3's 1B, 4B, and 12B, all deployed via Ollama with thinking mode disabled.
The format line directly determines how often the model shows its reasoning process: "Answer with the number only" led to 37% of responses showing reasoning, the plain version 27%, and the florid version only 8%.
The MoE architecture has been widely adopted in the large language model field in recent years, including in series such as Mixtral, DeepSeek, and Qwen3.
When running Qwen3 27B (MXFP4) at 16 concurrent sessions, aggregate prefill reaches 6,300–6,800 tok/s, and aggregate decode throughput peaks at approximately 900 tok/s.
The models hosted include the Qwen3 series (27B and the Next flash version) as well as DSV4 Flash.
Open-source models such as Mixtral, DeepSeek, and Qwen3 all adopt the MoE architecture.
40 more timeline events
All Facts (20)
The Qwen2.5:7B model is approximately 4 GB in size after Q4 quantization.
80%VerifiedQwen（通义千问）是阿里巴巴达摩院开源的系列大模型，覆盖从0.5B到72B的多种参数规模，支持多语言和多模态任务
80%VerifiedQwen3 introduces a dynamic switching mechanism between thinking mode and non-thinking mode.
80%VerifiedIn an RTX 5070 (12GB VRAM) configuration, the local Qwen3 9B model's output speed reached approximately 70–76 tokens/second.
70%VerifiedQwen3-series coding models load at approximately 20 GB and can expand to 35–40 GB during inference.
70%VerifiedQwen3 is a new generation open-source large model series released by Alibaba Cloud in 2025, supporting a hybrid reasoning mode.
70%VerifiedQwen (the Tongyi Qwen series) is a family of open-source large models launched by Alibaba.
65%VerifiedQwen3.8-27B is licensed under Apache 2.0, fully free for commercial use, with no additional commercial use restrictions.
65%VerifiedA 300B INT4 quantized model requires approximately 150–180GB of storage space.
65%VerifiedMoE models activate only a small fraction of total parameters per inference; Qwen3 Max may activate between 1/8 and 1/16 of the total parameters per inference.
65%VerifiedQwen系列采用Apache 2.0协议开放，允许商业使用和二次修改
65%VerifiedQwen3-4B is already on par with DeepSeek V3 or GPT-4o on multiple reasoning evaluation metrics.
65%VerifiedThe flagship version of Qwen3 has a parameter count as high as 235B and a size of approximately 142 GB.
65%VerifiedQwen3 can be deployed and run locally via Ollama.
65%VerifiedThe Qwen 2.5 series was launched in 2024, covering a full range of open-source weights from 0.5B to 72B.
75%UnverifiedRL训练中模型能够识别自己是否处于模拟环境中并据此改变行为，即奖励欺骗现象
90%UnverifiedGemma 4引入了多模态能力，Qwen3支持超长上下文，DeepSeek采用MoE（混合专家）架构
85%UnverifiedQwen3 8B版本在工具调用（Function Calling）基准测试中表现突出
70%UnverifiedMainstream open-source models such as Llama 3 and Qwen2/3 have already adopted GQA.
60%UnverifiedQwen3.5-4B is a small model with only 4 billion parameters, released by Alibaba's Tongyi Qianwen team.
60%