Search AI Products and News

The Tencent Hunyuan team announced the release of four open-source small-scale models with parameters of 0.5B, 1.8B, 4B, and 7B. These models are designed for consumer-grade graphics cards and are suitable for low-power scenarios such as laptops, smartphones, smart cockpits, and smart home devices. They support cost-effective fine-tuning in vertical domains. This move further enriches the Tencent Hunyuan open-source model system, providing developers and enterprises with more model size options. The release of these four models is part of Tencent Hunyuan's continuous open-source efforts, aiming to provide developers and enterprises with more choices to meet their needs.

The Qwen3-30B-A3B model has undergone a major update, launching the new version Qwen3-30B-A3B-Thinking-2507. This new version has achieved significant improvements in reasoning capabilities, general abilities, and context length, marking that the model is not only lighter but also has stronger thinking capabilities. Compared to the Qwen3-30-A3B model open-sourced on April 29, the new model has made progress in multiple key areas. In terms of mathematical abilities, Qwen3-30B-A3B-Thin

On July 29, a new version of the Qwen3-30B-A3B model, called Qwen3-30B-A3B-Instruct-2507, was released. This new version achieves significant improvements in multiple key areas, marking that the model can achieve performance comparable to top closed-source models such as Gemini2.5-Flash (non-thinking) and GPT-4o, by activating only 3B parameters in non-thinking mode.

Step Star releases the next generation Step 3 large model, which will be open-sourced on July 31st. This 321B parameter multi-modal reasoning model performs outstandingly in intelligence and reasoning efficiency, supporting cross-domain knowledge understanding and visual analysis. The company has established the Model Chip Ecosystem Innovation Alliance in cooperation with chip manufacturers and reached a strategic cooperation with Shanghai State Capital, promoting the application of large models. Step 3 achieves 300% reasoning efficiency of competing products on domestic chips, with over 70% improvement in throughput, and has established a 1+N model matrix covering multiple fields. The model has achieved SOTA on multiple rankings.

Kuaishou opensources the KAT-V1 autonomous thinking large model, which includes two versions: 40B and 200B. The 40B version performance is close to DeepSeek-R1, and the 200B version outperforms several flagship models. The model innovatively adopts a mixed training paradigm of short and long-term thinking and the Step-SRPO reinforcement learning algorithm, which can automatically adjust the thinking mode based on the complexity of the question, solving the problem of overthinking. Based on Qwen2.5-32B, it achieves excellent performance in fields such as science and code through an heterogeneous distillation framework and pre-training with 10 million examples.

AI office company PixelBloom completes B3 funding round, led by Yizhuang Guotou, with participation from multiple renowned institutions. The funds will be used for two strategic initiatives: accelerating global expansion and consolidating the leading position of AiPPT.com in international markets; expanding the AI Venture Studio model to incubate innovative office products. This funding round reflects market recognition of its technological capabilities and development prospects. In the future, it will promote the construction of the AI office ecosystem and enhance the intelligent office experience for global users.

On July 23, Kunlun Group introduced two innovative AI models to global users - the Mureka V7 Music Model and the Mureka TTS V1 Audio Model. The launch of these two models marks an important step forward in the field of AI music generation. Since its release, the Mureka V7 Music Model has attracted widespread attention, gaining extensive praise for its efficient music creation capabilities and outstanding music quality. Users simply need to input lyrics, select a style or theme, and click generate to obtain a complete and pleasant song in a short period of time.

ByteDance open sources the lightweight multilingual translation model Seed-X, which supports bidirectional translation in 28 languages and demonstrates performance comparable to top-level large models. This 7-billion-parameter model is based on the Mistral architecture and focuses on translation optimization, showing excellent performance in multiple areas. It uses an innovative training strategy to generate high-quality data and optimize deployment efficiency. This is another open-source project from ByteDance following BAGEL and Seed-Coder, promoting advancements in AI translation technology.

In 2025, open-source large models will show three major trends: 1) MoE architecture becomes mainstream, with DeepSeek-V3 (67.1 billion parameters) and Qwen3-235B (235 billion parameters) each having unique designs in their expert systems; 2) small models break through performance bottlenecks, with SmolLM3-3B adopting position-encoding-free technology, and Qwen3-4B achieving lightweight efficiency; 3) models show significant differentiation, with Llama3.2 focusing on general tasks, while Kimi-K2 (1 trillion parameters) excels in complex reasoning.

Lightricks releases open-source LTX-Video13B, a 13B-parameter video generation model with multi-scale rendering, achieving 30x faster speeds. It runs on consumer GPUs, supports 1216×704 real-time generation, and offers text/image/video-to-video modes. The model enhances coherence and detail, enabling keyframe control and style transfer. Free for SMEs, it includes training tools and optimized versions to democratize AI video creation.....

Moonshot AI's open-source model Kimi K2 excels in programming, matching Qwen3-235B-A22B and nearing o3-mini-high & Claude-3.7-Sonnet. With 1T MoE, 128k context, 65.8% accuracy on SWE-bench, and $0.14/M tokens, it's ideal for coding. Supports web generation, workflows, vLLM/Hugging Face deployment, MIT-licensed.....

NVIDIA's Canary-Qwen-2.5B sets a 5.63% WER record on Hugging Face OpenASR. This CC-BY licensed model combines FastConformer encoder with Qwen3-1.7B LLM decoder for efficient speech-to-text and NLP. Supports multi-GPU deployment for cloud/edge applications.....

Recently, the ByteDance Seed team collaborated with the University of Hong Kong and Fudan University to introduce an innovative reinforcement learning training method called POLARIS. This method successfully enhances the mathematical reasoning capabilities of small models to levels comparable to those of large models through a carefully designed Scaling RL strategy, offering a new approach for optimizing small models in the field of artificial intelligence. Experimental results show that the 4 billion parameter open-source model Qwen3-4B trained using POLARIS achieved remarkable performance on AIME25 and AIME24 mathematical tests.

Recently, Tencent Cloud officially launched the API service for the Tencent Hunyuan A13B model on its official website. The input price is set at 0.5 yuan per million Tokens, and the output price is 2 yuan per million Tokens, which has quickly sparked enthusiastic discussions in the developer community. As the first 13B-level MoE (Mixture of Experts) open-source hybrid inference model in the industry, Hunyuan-A13B features a total of 80B parameters and only 13B activated parameters, achieving performance comparable to leading open-source models of the same architecture, while also demonstrating efficient reasoning capabilities.

KeLing AI launched KeTu 2.1, enhancing instruction understanding, portrait beautification, cinematic effects, and 180+ style responses. It also improves text generation. Free for all members for 7 days, supporting text-to-image, single/multi-image reference.....

A major breakthrough has emerged in the AI video field — Vidu Q1 video model, launched by Shengshu Technology, officially released the Reference Generation feature, offering a revolutionary experience where video material can be generated from imagination in one step, redefining the technical boundaries and production efficiency of content creation. In traditional video production processes, creators have to go through complex steps such as script writing, character design, storyboard drawing, real scene shooting, and post-production editing, and the creation of a short film often takes several weeks or even months. The release of the Reference Generation feature in Vidu Q1 has completely broken this established model. Users