Qwen 3
The third-generation Qwen open-source large language model series released by Alibaba. It demonstrates strong performance in Chinese language understanding and generation, supports multiple parameter scales, and is available for local deployment.
Timeline (last 90 days)
The Qwen 3 series are text-only models, and images were not passed to them at all during testing; two of them fabricated the total amount of a receipt out of thin air.
Qwen 3 is an open-source multilingual large language model series from Alibaba.
The model uses a Mixture-of-Experts (MoE) architecture with 48 layers, 512 experts per layer, and 24,576 experts in total.
The model has a core of 125B parameters, plus an n-gram lookup table with 51B parameters and a multi-token prediction (MTP) draft layer with 4B parameters, totaling approximately 180B.
Strata is an inference engine published on GitHub, released by a developer for running the Qwen 3 series of large models.
The model has 512 experts per layer, but only 10 are activated per token, meaning the weights actually participating in computation are approximately 6B.
Unoptimized baseline token decoding speeds for running Qwen 3 (27B): RTX 5090 ~75, RTX 4090 ~46, RTX 3090 ~40, M5 Max 31.4, DGX Spark 12–17, Ryzen AI Max 11.5 tokens/s.
When using Hybrid Compute for the first time, three local model download options are offered: Gemma (suitable for low-end devices), Qwen 3 (approximately 35B parameters), and Perplexity's own model.
The Qwen 3 series is a multi-scale language model open-sourced by Alibaba Cloud, with its 8B/27B parameter versions capable of running locally on consumer-grade GPUs (such as the RTX 3090 with 24 GB VRAM)
Bonsai 2, released by PrismML, is a heavily quantized version based on Qwen 3.8 27B.
7 more timeline events
All Facts (14)
The model has 512 experts per layer, but only 10 are activated per token, meaning the weights actually participating in computation are approximately 6B.
60%UnverifiedThe Qwen 3 series are text-only models, and images were not passed to them at all during testing; two of them fabricated the total amount of a receipt out of thin air.
50%UnverifiedQwen 3 is an open-source multilingual large language model series from Alibaba.
50%UnverifiedThe model uses a Mixture-of-Experts (MoE) architecture with 48 layers, 512 experts per layer, and 24,576 experts in total.
50%UnverifiedThe model has a core of 125B parameters, plus an n-gram lookup table with 51B parameters and a multi-token prediction (MTP) draft layer with 4B parameters, totaling approximately 180B.
50%UnverifiedStrata is an inference engine published on GitHub, released by a developer for running the Qwen 3 series of large models.
50%UnverifiedUnoptimized baseline token decoding speeds for running Qwen 3 (27B): RTX 5090 ~75, RTX 4090 ~46, RTX 3090 ~40, M5 Max 31.4, DGX Spark 12–17, Ryzen AI Max 11.5 tokens/s.
50%UnverifiedWhen using Hybrid Compute for the first time, three local model download options are offered: Gemma (suitable for low-end devices), Qwen 3 (approximately 35B parameters), and Perplexity's own model.
50%UnverifiedThe Qwen 3 series is a multi-scale language model open-sourced by Alibaba Cloud, with its 8B/27B parameter versions capable of running locally on consumer-grade GPUs (such as the RTX 3090 with 24 GB VRAM)
50%UnverifiedQwen 3 is an open-source large language model series released by Alibaba's Tongyi Lab, featuring hybrid reasoning modes and reinforcement learning post-training capabilities.
50%UnverifiedThe tutorial recommends four mainstream models: Llama 3.1 (Meta) for the strongest overall capability, DeepSeek R1 (DeepSeek) for deep reasoning, Qwen 3 (Alibaba Tongyi Qianwen) for the best Chinese language performance, and Gemma 3 (Google) for multimodal support.
50%UnverifiedOLMo 2 made QK-Norm widely recognized as an important tool for training stability, and Gemma 3 and Qwen 3 adopted QK-Norm inspired by Meta Chameleon.
50%UnverifiedThe Qwen 3 series incorporates multiple architectural innovations, including hybrid reasoning.
50%UnverifiedBonsai 2, released by PrismML, is a heavily quantized version based on Qwen 3.8 27B.
50%