Jin10 data reported on July 9th that in the early hours of today, the globally renowned large model open source platform Hugging Face has open sourced the top small parameter model SmolLM3. SmolLM3 has only 3 billion parameters, yet its performance significantly surpasses similar open source models such as Llama-3.2-3B and Qwen2.5-3B. It features a 128k context window and supports six languages including English, French, Spanish, and German. It supports both depth thinking and non-thinking dual reasoning modes, allowing users to switch flexibly.

QwQ-Max will soon be released in general availability, with Open Source QwQ-Max and Qwen 2.5-Max under the Apache 2.0 license, and smaller versions such as QwQ-32B, which can be deployed on local devices. At the same time, the official QwQ-Max will be released, and Android and iOS apps will be released.

Jinshi Data News on February 25th, the Alibaba Qwen team announced the release of a new inference model on social media this morning - Depth Thinking (QwQ). This is an inference model based on Qwen2.5-Max supported by QWQ-MAX-PREVIEW. Blue Whale News found that similar DeepSeek

Gate.io News

Jingwei Hengrun uses large models such as DeepSeek and Qwen to develop professional AI applications, covering multiple fields such as intelligent assistance, design, testing, product optimization, and business process intelligence. These applications improve work efficiency, quality, and reduce costs.

On February 26, Jinshi data news, it was learned from the Fourth Paradigm that the Fourth Paradigm launched a large model inference edge solution ModelHub AIoT. Users can easily deploy small distillation models including DeepSeek R1, Qwen 2.5, and Llama 2/3 series at the edge, and achieve offline operation. Users can flexibly switch between multiple models, taking into account model compression, inference performance, and solving the complexity of deployment and optimization. The company stated that this solution not only meets users' needs for privacy and real-time performance but also greatly reduces the cost of AI large model inference.

The Fourth Paradigm launches the large model inference integrated machine SageOneIA, reducing the cost of large model inference. Intelligent Computing Power pooling technology is integrated into the integrated solution, supporting multiple mainstream large models. Enterprises can switch between full-blooded and distilled models, with GPU utilization rate increased by more than 30% and inference performance improved by 5-10 times.

On February 12, Golden Ten Data reported that Ali International Station AI has accessed DeepSeek and other large models, which will be fully applied to various core aspects of foreign trade business. After accessing, AI will be able to comprehensively consider the buyer's inquiries when automatically receiving overseas customers, and then combine market insights, foreign trade experience, and the seller's own product situation on Ali International Station for Depth reasoning. It is also known that Ali International Station is testing universal Qwen questions.

Microsoft has released the latest version V2.0 of OmniParser, improving the efficiency and accuracy of the AI Agent used on computers, especially in detecting small UI elements and improving inference speed. The performance of this version in the high-resolution AgentBenchmark test has also been greatly improved, reaching an amazing 39.6% accuracy. In addition, Microsoft has Open Sourced the key tool omnitool for converting large models into Agents.