News from the AI & ML world
DeeperML - #deepseekv3
|
Nehdiii@Towards AI
//
DeepSeek AI has released its V3-0324 endpoint, offering AI developers access to a powerful 685 billion parameter model. This new endpoint boasts lightning-fast responses and a massive 128K context window, accessible via a simple API key. The model is available without rate limiting at a cost-effective price of $0.88 per 164K output tokens, making it an attractive option for developers seeking high performance at a reasonable price.
Lambda is offering DeepSeek V3-0324 live on its Inference API, providing developers with easy access to this powerful AI model. Towards AI has published a series of articles on DeepSeek V3, including a piece on auxiliary-loss-free load balancing. The DeepSeek V3-0324 highlights includes Major Boost in Reasoning performance, 685B total parameters using a Mixture-of-Experts (MoE) design, Stronger front-end development skills and Smarter tool-use capabilities. However, DeepSeek faces competition from other AI companies, particularly China's Baidu. Baidu recently launched two new AI models, ERNIE X1 and ERNIE 4.5, aiming to compete in the global race for advanced AI. According to TheTechBasic, ERNIE X1 is designed to match DeepSeek R1 in performance but at half the price, while ERNIE 4.5 is capable of handling text, video, images, and audio with improved logic and memory skills. Baidu hopes these new models will help it regain ground against rivals.
Share:
References :
Classification:
@syncedreview.com
//
DeepSeek AI is making waves in the large language model (LLM) field with its innovative approach to scaling inference and its commitment to open-source development. The company recently published a research paper detailing a new technique to enhance the scalability of general reward models (GRMs) during the inference phase. This new method allows GRMs to optimize reward generation by dynamically producing principles and critiques, achieved through rejection fine-tuning and rule-based online reinforcement learning. Simultaneously, DeepSeek AI has hinted at the imminent arrival of its next-generation model, R2, sparking considerable excitement within the AI community.
DeepSeek’s advancements come at a crucial time, as the focus in LLM scaling shifts from pre-training to post-training, especially the inference phase. The company's R1 series already demonstrated the potential of pure reinforcement learning in enhancing LLM reasoning capabilities. Reinforcement learning serves as a vital complement to the "next token prediction" mechanism of LLMs, providing them with an "Internal World Model." This enables LLMs to simulate different reasoning paths, evaluate their quality, and choose superior solutions, ultimately leading to more systematic long-term planning, the company, in collaboration with Tsinghua University, unveiled a new research study aimed at improving reward modelling in large language models by utilising more inference time compute. This research resulted in a model named DeepSeek-GRM, which the company asserts will be released as open source. Further emphasizing its dedication to accessibility and collaboration, DeepSeek AI is planning to open-source its inference engine. DeepSeek AI’s dedication to open-sourcing key components and libraries of its models. The company is "collaborating closely" with existing open-source projects and frameworks to ensure seamless integration and widespread adoption. Additionally, DeepSeek released five high-performance AI infrastructure tools as open-source libraries during Open Source Week, enhancing the scalability, deployment, and efficiency of training large language models. DeepSeek’s efforts reflect a broader industry trend towards leveraging open-source initiatives to accelerate innovation and democratize access to advanced AI technologies.
Share:
References :
Classification:
|
BenchmarksBlogsResearch Tools |