Alibaba has launched its Qwen3.5 Small Model Series, featuring compact, open-source AI models that are making significant waves in the tech world. Notably, the Qwen3.5-9B model is proving to be a formidable competitor, outperforming OpenAI's much larger gpt-oss-120B on key benchmarks, including graduate-level reasoning and multilingual knowledge. This development is particularly impressive considering the Qwen3.5-9B is over 13 times smaller than its OpenAI counterpart.
The Qwen3.5 series includes several models, each optimised for different use cases. The 0.8B and 2B models are designed for speed and efficiency on edge devices, perfect for prototyping, while the 4B model serves as a robust multimodal base with a substantial context window. The star of the show, the 9B model, demonstrates remarkable reasoning capabilities and multimodal understanding, such as interpreting UI elements and counting objects in videos. This native multimodality is achieved through an "early fusion" training approach, a departure from traditional methods.
Technically, these models employ an Efficient Hybrid Architecture, combining Gated Delta Networks with sparse Mixture-of-Experts. This innovative design addresses the limitations often faced by smaller models, resulting in higher throughput and lower latency. The weights for all Qwen3.5 models are globally available under Apache 2.0 licenses on Hugging Face and ModelScope, encouraging enterprise and commercial use, including customisation. The performance benchmarks are striking, with the Qwen3.5-9B surpassing competitors like Gemini 2.5 Flash-Lite in visual reasoning and exhibiting strong capabilities in video understanding and mathematical problem-solving. The open-source nature and impressive efficiency of these models are enabling developers to run sophisticated AI tasks on standard laptops and even mobile phones, democratising access to advanced AI capabilities and paving the way for the next generation of autonomous agents.
Fuente Original: https://venturebeat.com/technology/alibabas-small-open-source-qwen3-5-9b-beats-openais-gpt-oss-120b-and-can-run
Artículos relacionados de LaRebelión:
- OpenAIs Audio AI Voice Models Hardware Coming
- OpenAIs 1 Trillion IPO Worth The Risk
- DeepSeeks AI Model 294K Training Cost
- Huawei Chips Stall DeepSeek R2 AI Model A Tech Setback
- AI2s MolmoAct A 3D Robotics AI Model Challenging Nvidia and Google
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario