Marktechpost MTP: the Mathstral model, BigVGAN v2, DeepSeek-V2-0628 and Deepset-Mxbai-Embed-de-Large-v1 Released.....

Featured Research

Mistral AI announces the release of its latest model, the Mathstral model. This new model is specifically designed for mathematical reasoning and scientific discovery. Named as a tribute to Archimedes, whose 2311th anniversary is celebrated this year, Mathstral is a 7-billion parameter model with a 32,000-token context window, published under the Apache 2.0 license.

Mathstral is introduced as part of Mistral AI’s broader effort to support academic projects developed in collaboration with Project Numina. This new model aims to bolster efforts in tackling advanced mathematical problems requiring complex, multi-step logical reasoning. It is akin to Isaac Newton standing on the shoulders of giants, building upon the capabilities of the Mistral 7B model and specializing in STEM (Science, Technology, Engineering, and Mathematics) subjects. Mathstral achieves state-of-the-art reasoning capacities in its size category across various industry-standard benchmarks, scoring 56.6% on MATH and 63.47% on MMLU.

 Editor’s Picks…

Nvidia AI Releases BigVGAN v2: A State-of-the-Art Neural Vocoder Transforming Audio Synthesis

In the rapidly developing field of audio synthesis, Nvidia has recently introduced BigVGAN v2. This neural vocoder breaks previous records for audio creation speed, quality, and adaptability by converting Mel spectrograms into high-fidelity waveforms. This team has thoroughly examined the main enhancements and ideas that set BigVGAN v2 apart.

One of BigVGAN v2’s most notable features is its unique inference CUDA kernel, which combines fused upsampling and activation processes. With this breakthrough, performance has been greatly increased, with Nvidia’s A100 GPUs attaining up to three times faster inference speeds. BigVGAN v2 assures that high-quality audio may be synthesized more efficiently than ever before by streamlining the processing pipeline, which makes it an invaluable tool for real-time applications and massive audio projects.

ADVERTISEMENT

[Synthetic Data Webinar] Learn how Gretel’s synthetic data platform, powered by generative AI, make’s data generation easier than ever before..

During this webinar, you will see live demos of the Gretel platform and learn about the latest product additions:

🐝 Gretel Navigator: Our new agent-based, compound AI system tailor-made for tabular data generation

🐝 Gretel Open Datasets: We’ve released a few open source datasets including the world’s largest text-to-SQL dataset

🐝 Navigator Fine Tuning: Fine-tune a specialized language model on your unique, domain-specific data

🐝 Transform v2: Apply flexible de-identification and rule-based transformations to real and synthetic datasets

and many more…

DeepSeek-V2-0628 Released: An Improved Open-Source Version of DeepSeek-V2

DeepSeek-V2-Chat-0628 is an enhanced iteration of the previous DeepSeek-V2-Chat model. This new version has been meticulously refined to deliver superior performance across various benchmarks. According to the LMSYS Chatbot Arena Leaderboard, DeepSeek-V2-Chat-0628 has secured an impressive overall ranking of #11, outperforming all other open-source models. This achievement underscores DeepSeek’s commitment to advancing the field of artificial intelligence and providing top-tier solutions for conversational AI applications.

The improvements in DeepSeek-V2-Chat-0628 are extensive, covering various critical aspects of the model’s functionality. Notably, the model exhibits substantial enhancements in several benchmark tests

The DeepSeek-V2-Chat-0628 model also features optimized instruction-following capabilities within the “system” area, significantly enhancing the user experience. This optimization benefits tasks such as immersive translation and Retrieval-Augmented Generation (RAG), providing users with a more intuitive and efficient interaction with the AI.......

Deepset-Mxbai-Embed-de-Large-v1 Released: A New Open Source German/English Embedding Model

Deepset and Mixedbread have taken a bold step toward addressing the imbalance in the AI landscape that predominantly favors English-speaking markets. They have introduced a groundbreaking open-source German/English embedding model, deepset-mxbai-embed-de-large-v1, to enhance multilingual capabilities in natural language processing (NLP).

This model is based on intfloat/multilingual-e5-large and has undergone fine-tuning on over 30 million pairs of German data, specifically tailored for retrieval tasks. One of the key metrics used to evaluate retrieval tasks is NDCG@10, which measures the accuracy of ranking results compared to an ideally ordered list. Deepset-mxbai-embed-de-large-v1 has set a new standard for open-source German embedding models, competing favorably with commercial alternatives.

Other Trending AI Updates

• Nephilim v3 8B Released: An Innovative AI Approach to Merging Models for Enhanced Roleplay and Creativity

• ZebraLogic: A Logical Reasoning AI Benchmark Designed for Evaluating LLMs with Logic Puzzles

• Mistral AI and NVIDIA Collaborate to Release Mistral NeMo: A 12B Open Language Model Featuring 128k Context Window, Multilingual Capabilities, and Tekken Tokenizer

• Hugging Face Introduces SmolLM: Transforming On-Device AI with High-Performance Small Language Models from 135M to 1.7B Parameters

• This AI Paper from NYU and Meta Introduces Neural Optimal Transport with Lagrangian Costs: Efficient Modeling of Complex Transport Dynamics

Upcoming AI Webinars (July 22-31, 2024)

Here is a list of Upcoming AI Webinars (July 21- 31, 2024) from various AI and Data Companies