Researchers at Johns Hopkins University and Tencent AI Lab have developed a new text-to-audio model called EzAudio. As Michael Nuñez reports for VentureBeat, EzAudio can generate high-quality sound effects from text descriptions. The model uses an innovative method for processing audio data and a new architecture called EzAudio-DiT. In tests, EzAudio outperformed existing open-source models …
Companies can achieve significant cost savings by building their own conversational AI based on the open source model Llama 3. This is the result of an analysis by Sam Oliver, founder of OpenFi, published on VentureBeat. According to Oliver, an average conversation with Llama 3 costs about $0.08, while the same conversation with OpenAI’s GPT-4 …
DeepSeek-V2.5 is the new champion among open source AI models. DeepSeek itself is an offshoot of the Chinese hedge fund High-Flyer Capital Management. The new model combines natural language processing and programming capabilities into one seemingly powerful system. According to Carl Franzen of VentureBeat, DeepSeek-V2.5 outperforms its predecessors in almost all benchmarks. The model offers …
A new open source model called OLMoE has been released by the Allen Institute for AI (AI2) in collaboration with Contextual AI. As Emilia David reports for VentureBeat, the model aims to be both powerful and inexpensive. OLMoE uses a mixed-expert architecture with 7 billion parameters, of which only 1 billion are active per input …
What a difference a weekend makes: Just a few days ago, the new Reflection 70B was considered the potential champion among open source models. According to official information, the model is based on Meta’s Llama 3.1 70B, but with a special feature: it is specially trained not to give an immediate answer, but to analyze …
A new open source AI model called Reflection 70B has been introduced by Matt Shumer, co-founder of AI startup HyperWrite. As Shumer announced on the platform X (formerly Twitter), the model outperforms leading commercial systems in benchmarks. Reflection 70B is based on Meta’s Llama 3.1-70B Instruct and uses a new technique for self-correcting errors: the …
A new AI application called AnythingLLM allows users to chat with any document. The software supports several AI language models and vector databases. According to the developers, users can use it to create a private ChatGPT-like application that can be hosted locally or remotely. AnythingLLM offers features such as multimodality, multi-user support and embedded chat …
Hugging Face’s new ranking of the best freely avaiable language models shows that Chinese models currently lead the way. Alibaba’s Qwen models dominate the top spots in the ranking, which is based on more challenging tests than its predecessor. Skills such as knowledge recall, inferring from long texts, complex mathematics, and following instructions are assessed.
Many AI models that power chatbots advertise themselves as “open source,” but do not fully release the code and training data. A new study shows that many large companies describe their models as “open weights”, meaning that researchers can use them, but have no access to the underlying data and can’t make fundamental changes to …
Stability AI has released Stable Diffusion 3 Medium, a smaller version of its image generation model that can run on PCs with as little as 5GB of VRAM. According to Stability AI, the model offers comparable quality to the larger version and could therefore be an attractive option for users with limited resources.