Last Week in AI: Announcements from Timescale, Fixie AI, Nexusflow, and More -- Pure AI

Last Week in AI: Announcements from Timescale, Fixie AI, Nexusflow, and More

By Pure AI Editors
11/18/2024

This edition of our roundup of AI products and services announced last week that didn't make splashy headlines, but should be on your radar, includes Timescale's Vectorizer, Fixie AI's Ultravox v0.4.1, Nexusflow's Athene-V2, Ironclad's Jurist, and more.

Timescale, a leading PostgreSQL database platform, unveiled pgai Vectorizer, an open-source and cloud-hosted tool that integrates advanced AI capabilities directly into PostgreSQL, the world’s most popular database. This integration is meant to empower developers to build AI applications seamlessly, eliminating the need for specialized expertise or external tools. The release, part of Timescale’s pgai suite, which includes pgai and pgvectorscale, enables developers to create, manage, and synchronize vector embeddings within PostgreSQL. This simplifies the AI development process, reducing infrastructure costs by 75% and making AI accessible to the 49% of developers worldwide who already rely on PostgreSQL, the company says.

Key features of pgai Vectorizer include:

Unified AI and Data Platform: Manage vector embeddings, metadata, and event data within PostgreSQL.
Real-Time Synchronization: Ensure up-to-date embeddings synchronized with data changes.
Seamless Experimentation: Switch embedding models effortlessly without altering application code.
Version Tracking: Enable smooth rollouts with backward compatibility for model updates.

Fixie AI announced the release of Ultravox v0.4.1, a cutting-edge, open-source model designed to enhance real-time conversational AI. Ultravox v0.4.1 supports multi-modal inputs, including text, images, and other sensory data, and offers faster, more fluid interactions, providing an alternative to proprietary models like GPT-4, the company says. Built on a transformer-based architecture with cross-modal attention, Ultravox was designed to enable seamless integration of diverse data formats for applications ranging from customer support to education. The models, hosted on Hugging Face, are open to developers and researchers worldwide, facilitating innovation and transparency in conversational AI. Boasting a 30% reduction in response latency compared to commercial counterparts, Ultravox delivers high accuracy and context-aware dialogue capabilities. Its adaptability positions it as a valuable tool for industries requiring real-time, multi-modal AI, such as healthcare and interactive learning, the company says.

Nexusflow announced Athene-V2, a new suite of fine-tuned 72B AI models designed to compete with GPT-4o across specialized use cases. Built on Qwen 2.5 72B and enhanced through advanced post-training and reinforcement learning (RLHF) pipelines, Athene-V2 marks a strategic shift from universal AI capabilities toward targeted customization, the company says.

The suite includes two models optimized for distinct roles:

Athene-V2-Chat-72B: Excels in dialogue tasks, surpassing GPT-4o in chat helpfulness (Arena-Hard) and ranking #2 on bigcode-bench-hard for coding. It also demonstrates superior performance in mathematics (MATH) and long log extractions.
Athene-V2-Agent-72B: Balances chat and agent capabilities, delivering concise responses and outperforming GPT-4o in enterprise-level function calling benchmarks.

Nexusflow highlights a "Pareto frontier" approach, focusing on refining specific model capabilities rather than universal performance gains. This strategy enables Athene-V2 to deliver advanced precision-recall balancing, critical for complex real-world applications like ticket management systems, the company says.

Ironclad, a leading digital contracting platform, announced the public release of Ironclad Jurist, a conversational AI assistant designed for legal workflows. Jurist allows legal professionals to draft, edit, review, summarize, and translate legal documents while providing real-time insights and benchmarks, all within a fully editable online .docx workspace, the company says. Built on Ironclad's open-source Rivet platform, Jurist employs such AI tech as prompt routing, legal-specific prompt engineering, and a retrieval automation generation (RAG) approach. The assistant ensures transparency by displaying agent reasoning and citations for its actions, giving legal teams confidence in its outputs, the company says.

Key features of Jurist include:

Centralized Legal Workflows: Lawyers can draft, edit, research, and ask questions about contracts in one environment.
Personalized Outputs: The assistant tailors drafts based on templates and past agreements provided by users.
Up-to-Date Legal Knowledge: Jurist leverages verified online sources to stay current with the latest legal developments.
Responsible Data Practices: Ironclad ensures privacy with enterprise-grade security, GDPR compliance, and certifications like SOC 2 Type II.

Qwen open sourced its Qwen2.5-Coder series, a family of advanced coding language models designed to enhance the efficiency and accuracy of coding tasks. Built on the Qwen2.5 architecture, the series includes models ranging in size from 0.5 billion to 32 billion parameters, offering flexibility for developers, researchers, and industry professionals. The flagship Qwen2.5-Coder-32B-Instruct model delivers state-of-the-art performance, excelling in benchmarks such as HumanEval and BigCodeBench and surpassing competitors in accuracy and multi-language coding capabilities. Pretrained on over 5.5 trillion tokens and fine-tuned to ensure high-quality, executable outputs, the models support a wide range of applications, including code generation, completion, and reasoning. Qwen's open-source release underscores its commitment to innovation and accessibility, making scalable and versatile coding tools available to a broader community, the company says.

Eviden, a division of the Atos Group, announced the launch of BXI v3, the third generation of its BullSequana eXascale Interconnect technology. The new scale-out networking solution, designed for AI and high-performance computing (HPC) workloads, is set for release in the second half of 2025. BXI v3 addresses the growing performance gap between compute power and networking in exascale workloads and large language models. Developed in partnership with France’s Atomic Energy Commission (CEA), the technology features SmartNIC capabilities that offload application communications from processors, boosting efficiency. The system leverages Ethernet as its base protocol, enhanced with advanced features like low latency, high bandwidth, and congestion management. Eviden says BXI v3 improves application performance by up to 35% while lowering total cost of ownership. As a founding member of the Ultra Ethernet Consortium, Eviden aims to extend Ethernet’s capabilities for AI and HPC, fostering broader industry collaboration, the company says.

Alif Semiconductor and Edge Impulse announced what they are calling a breakthrough in AI vision processing for edge microcontrollers with full integration of Nvidia's TAO model training toolkit into Alif's Ensemble and Balletto MCU families via the Edge Impulse platform. This marks the first proven deployment of TAO-trained models on low-power edge devices, they companies said. The TAO toolkit, known for its comprehensive datasets and transfer learning capabilities, simplifies the development of AI vision applications like people counting and intruder detection. Alif's MCUs, featuring Arm Ethos-U55 neural processing units optimized for TAO, now support streamlined workflows for deploying pre-trained or custom models using Edge Impulse.