Generative AI

Harvey has released Harvey Tenet, its first post-trained model, as a research preview on August 20, 2026. Tenet is built upon the Kimi K3 base model and has undergone post-training using Fireworks through asynchronous reinforcement learning, specifically targeting long-horizon legal work. The comprehensive training corpus integrated synthetic data, publicly accessible legal datasets, and data provided by human legal experts. Harvey has explicitly stated that no customer data was utilized in the training process. When evaluated against the base K3 model, Tenet demonstrated a significant improvement, completing nearly twice as many held-out tasks on Harvey's proprietary Legal Agent Benchmark (LAB). Furthermore, it achieved 20% more completions on the LAB: Contracts sub-benchmark, leading to an increase in the overall pass rate by 9 percentage points and the contracts pass rate by 2 percentage points. Harvey reports that Tenet has achieved state-of-the-art performance on LAB: Contracts and secured second place on the broader LAB. These performance gains have also shown transferability to other agent systems, including Mercor's APEX Agents and Crosby's Redline Bench, even without further training on those platforms. The dual objectives behind the development of Harvey Tenet are to advance frontier legal intelligence using open-weight models and to provide law firms with the capability to develop and own their specialized AI models. However, Harvey Tenet is not yet deployable as a standalone product. As of its announcement on August 20, 2026, it remains a research preview. Harvey has not yet published the model's weights, a detailed model card, or an API endpoint for external access. While the base Kimi K3 model is open-weight, Tenet itself is a proprietary checkpoint developed by Harvey. The company indicates that the advancements made in this research will be integrated into Harvey's existing products over time, transitioning from research to production. The current offering provides the methodology and training approach rather than a ready-to-use artifact. Access to Harvey's platform, which is sold to law firms, mid-sized firms, and in-house legal teams, is required for enterprise-tier engagement. A laboratory equipped with a reinforcement learning stack could potentially replicate the training methodology. The training process involved approximately 150 NVIDIA B300 GPUs utilized over a two-month period. The primary industries targeted for this technology include legal services, corporate in-house legal departments, private equity and investment banking for M&A diligence, and regulated sectors with high contract volumes such as insurance, financial services, healthcare, and energy. Potential applications encompass the generation of M&A due diligence memos from datarooms, contract drafting, review, and redlining, structured data extraction from up to 10,000 documents, and precedent search across a firm's internal knowledge base.

The rapid advancement of open-weight large language models (LLMs) has outpaced the hardware assumptions traditionally associated with their deployment. Models such as Kimi-K3, GLM-5.2, and DeepSeek-V4-Flash are rapidly closing the capability gap with proprietary AI systems. However, merely releasing model parameters does not address the significant challenge of affordability and accessibility for running these powerful tools. Historically, serving LLMs has necessitated extensive datacenter-class GPU clusters, a prohibitive cost for individual developers and smaller teams, especially as agentic workloads increasingly drive inference demand. This landscape is shifting with the introduction of FreeToken, an edge-native Mixture-of-Experts (MoE) serving engine developed by researchers from the University of California, Berkeley, and the University of Texas at Austin. FreeToken redefines the concept of an inference platform by treating a personal machine as a unified, elastic resource. Instead of being limited by a single GPU's capacity, it intelligently and continuously maps computation and model state across all available hardware components, including the GPU, CPU, system memory, and interconnect bandwidth. This dynamic allocation allows for unprecedented performance on consumer-grade hardware. For instance, a 35 billion parameter model can achieve interactive speeds on a laptop with just 8GB of GPU memory, a 284 billion parameter model can run on a standard gaming desktop, and remarkably, the colossal 753 billion parameter GLM-5.2 model can be served on a single workstation GPU. FreeToken is designed for broad deployability and accessibility. It is open-sourced under the permissive Apache-2.0 license on GitHub and is readily available on the Python Package Index (PyPI) as `freetoken v0.1.2`, installable via `pip install "freetoken[accel]"`. Furthermore, a user-friendly one-click desktop application for both Windows and Linux is distributed through flashml.ai. The command-line interface (CLI) is currently optimized for Linux x86_64 systems equipped with NVIDIA GPUs running driver r580 or later (supporting CUDA 13). The `ft serve` command within FreeToken exposes OpenAI- and Anthropic-compatible API endpoints on port 1919, enabling seamless integration with existing workflows. Additionally, the `ft launch claude` command simplifies the process of wiring up powerful models like Claude Code, Codex, OpenCode, or OpenClaw directly to a user's local machine. FreeToken is particularly well-suited for solo developers, startups, and small to medium-sized business (SMB) engineering teams whose escalating token bills for AI agents are beginning to outweigh the cost of owning dedicated GPU hardware. For larger enterprises, FreeToken presents a strategic solution for implementing air-gapped or regulated workloads, offering a secure, on-premise inference path rather than a direct replacement for established datacenter infrastructure. Industries that stand to benefit most include healthcare and legal sectors, where maintaining strict data privacy and ensuring data never leaves the local machine is paramount. Other strong fits include defense, finance, and intellectual property-intensive research and development environments. Typical applications leveraging FreeToken include the development of local coding agents, private code review tools, offline contract analysis, and the generation of synthetic data, all processed locally without any sensitive information being transmitted externally. This innovation directly addresses the growing disparity between the rapid proliferation of highly capable open-weight AI models and the practical, affordable means for their deployment and utilization by a wider range of users and organizations.

DeepDoctection version 1.2.x facilitates the creation of end-to-end document intelligence pipelines, combining multiple processing stages into a single, cohesive workflow. This tutorial demonstrates how to configure an analyzer explicitly for tasks including layout detection, table structure recognition, optical character recognition (OCR), reading-order reconstruction, annotation linking, and structured data export. The framework represents extracted information through Page objects, detailing text, figures, tables, their interrelationships, provenance, and the sequence in which content should be read. The implementation utilizes specific components within the deepDoctection ecosystem. For layout detection, the pipeline is configured with models based on the DocLayNet dataset. Table structure recognition is handled by the Table Transformer model, and OCR is performed using the DocTR library. Users can inspect the resulting Page objects to understand the detailed representation of document content and its structure. The framework is designed to be extensible, allowing developers to register custom object types and implement their own PipelineComponents. This customizability is showcased through the development of a component for extracting monetary and date entities, alongside classifying documents based on their tabular characteristics. Further customization involves assembling a custom pipeline manually using the ServiceFactory, which provides granular control over the processing flow. This includes exploring options for filtering intermediate results and implementing service rollback mechanisms. The processed page data can then be serialized, and document annotations are transformed into ordered JSONL (JSON Lines) chunks. These structured outputs are specifically formatted to be suitable for downstream applications such as Retrieval Augmented Generation (RAG) systems and other retrieval-based AI models, enabling more efficient and accurate information retrieval from documents. The tutorial also details the installation process for necessary libraries, including deepdoctection itself, transformers, timm, python-doctr, pdfplumber, networkx, and lxml. Environment variables are set to configure the deepdoctection environment, such as enabling PyTorch, setting the DPI for image processing to 200, and adjusting the log level to INFO. A specific patch is applied to disable PEFT adapter lookup for `from_pretrained` within the transformers integration, ensuring compatibility and preventing potential issues. Sample documents, including a PDF file named 'paper.pdf' and an image file 'finance.png', are downloaded to facilitate practical demonstration of the pipeline's capabilities.

NVIDIA has released a tutorial detailing how to build an advanced NeMo Guardrails pipeline designed to enhance enterprise AI safety by controlling Large Language Model (LLM)-based applications throughout the entire request lifecycle. The tutorial focuses on constructing an LLM-based financial assistant that integrates multiple layers of security and control mechanisms. This approach aims to ensure that AI systems operate within defined boundaries, preventing unintended or harmful outputs and maintaining compliance with organizational policies. The core of the tutorial involves combining deterministic Personal Identifiable Information (PII) detection and redaction with LLM-based self-checks for both input and output. Retrieval filtering is also implemented to manage the information the AI can access, alongside specific controls like account number masking and topical restrictions. Furthermore, the system incorporates policy-based tool gating, which dictates which tools or functions the AI can utilize based on predefined rules. This comprehensive setup allows for stateful multi-turn interactions, meaning the AI can maintain context and engage in more natural, extended conversations while adhering to safety protocols. To ensure transparency and evaluability, the pipeline includes detailed tracing of rail activations, enabling developers to understand precisely which control mechanism handled each request. Token accounting is also implemented to monitor computational costs associated with these safety features, providing insights into resource utilization. A key component of the evaluation process is a red-team-style coverage report. This report assesses whether the financial assistant responds safely and quantifies the computational expense incurred by the implemented protections. The tutorial utilizes OpenAI's GPT-4o-mini model as the LLM engine, configurable via an API key, and specifies a base URL for the model's endpoint if not using the default OpenAI service. The configuration file, written in YAML, outlines the model to be used and defines the various guardrails. The input rails include PII redaction and an input self-check. The retrieval process is filtered, and the output rails feature account number masking and an output self-check. The self-check prompt for input is designed to block user messages that attempt to override instructions, impersonate unrestricted assistants, contain abusive language, or try to access other customers' accounts, while allowing for ordinary complaints or small talk. This detailed framework provided by NVIDIA aims to equip developers with the tools to build robust and secure AI applications for enterprise use, addressing critical concerns around data privacy, operational integrity, and user safety.