KPMG, one of the world's largest professional services firms, announces a global alliance with Anthropic to embed Claude into its Digital Gateway platform for tax and legal clients, and provide access to over 276,000 employees worldwide. Anthropic names KPMG a preferred partner for private equity, jointly developing Claude-powered products for portfolio companies.

Anthropic is organizing dialogues with wisdom traditions—including scholars, clergy, philosophers, and ethicists from over 15 religious and cross-cultural groups—to inform AI development. These conversations aim to bring diverse perspectives to Claude's values, behaviors, and constitution, and have already inspired experimental ideas like a tool that gives Claude an ethical reminder to reduce misaligned behavior. Future engagements will include legal scholars, psychologists, writers, and civic institutions.

ElevenLabs launches two education initiatives: an access program for professors to use voice AI in classrooms, and an Einstein voice agent for interactive learning experiences.

Cerebras launches enterprise trials of Kimi K2.6, a trillion-parameter open-weight model, achieving 981 tokens per second inference speed—6.7x faster than GPU cloud. The model excels at coding and agentic tasks, enabling real-time development productivity boost.

This week's LlamaIndex newsletter highlights ParseBench, the first OCR benchmark for AI agents, along with new open-source tools: a sandboxed CLI agent for secure document interactions and a self-hostable document parsing server. Community events in Singapore and NYC are also recapped.

Anthropic acquires Stainless, a leader in SDKs and MCP server tooling, to extend Claude's ability to connect to data and tools. Stainless has powered Anthropic's official SDKs since the early days, serving hundreds of companies. The acquisition aims to push the frontier of developer experience and agent connectivity.

Baseten optimized image generation for Flux.2 [dev] and Qwen-Image, achieving up to 2.3x and 1.6x speedups on NVIDIA Blackwell GPUs, and significant gains on Hopper GPUs, using quantization, optimized kernels, and runtime improvements.

Cursor launches Composer 2.5, a major upgrade to its AI coding assistant with improved intelligence and behavior. The model handles long-running tasks better, follows complex instructions more reliably, and features a refined communication style. Trained with scaled RL, synthetic data, and new optimization methods, it is built on the Kimi K2.5 checkpoint. Pricing starts at $0.50/M input tokens and $2.50/M output tokens, with a faster variant at $3.00/M input and $15.00/M output. Usage is doubled for the first week.

This post walks through a demo AI agent that ingests SEC filings, searches across them, and answers questions with precise citations that highlight the exact source text on the original PDF page. The key ingredient is LiteParse, which extracts text along with bounding box coordinates. The project uses a simple keyword search instead of vector databases, and integrates SEC EDGAR for direct filing retrieval.

The Naval Sea Systems Command awarded Unstructured a contract to design an AI-enabled solution that helps warfighters surface mission-critical information faster, reduce operator workload, and accelerate decision-making in Anti-Submarine Warfare and surface warfare operations. The solution integrates Unstructured's data ingestion with Elastic's enterprise search to mine heterogeneous data sources, initially deployed on CV-TSC and USW-DSS systems with future applicability to JADC2 and C5ISR.

Unstructured introduces Extract, a new enrichment node that extracts structured JSON data from documents using LLM or regex, enabling intelligent document processing within existing workflows.

Unstructured launches webhooks to automate downstream actions based on job lifecycle events, allowing integration with any endpoint via workspace or workflow scopes.

Unstructured found that combining high-quality datasets with incompatible annotation styles degraded model performance. They built an agentic harmonization pipeline using a VLM to reconcile label differences, resulting in improved metrics across 14 of 17 benchmarks.

Unstructured's SCORE-Bench benchmark evaluates five frontier models on enterprise document parsing, revealing a significant gap between raw model calls and optimized pipelines. While models excel in reasoning and hallucination control (especially Claude Opus 4.6), they lag by up to 23 percentage points in table extraction, document structure, and output consistency. The gap is attributed to configuration rather than capability, and can be closed via optimized prompting, post-processing, and output structure enforcement.

Unstructured's new guide on advanced Retrieval-Augmented Generation (RAG) techniques, covering smart chunking, metadata filtering, GraphRAG, hybrid search, and agentic workflows, aimed at building scalable enterprise AI pipelines.

Unstructured announces major updates including a simplified drag-and-drop interface, Generative Refinement for higher fidelity outputs, and simplified pricing with a free tier. The new workflow combines high-resolution partitioning with VLM-powered enrichments to achieve superior accuracy and structure preservation.

Most teams still evaluate LLMs using tokens per second and cost per million tokens, but these metrics fail to predict production behavior. This article reveals the real trade-offs among speed, cost, and quality, introduces the Pareto frontier as an evaluation framework, and highlights critical production metrics like TTFT and p99 latency.

This guide details six production-tested optimization strategies for LLM inference, helping teams match specific bottlenecks with the highest-impact methods, including batching, prefill/decode optimizations, KV cache optimizations, attention/memory optimizations, parallelism, and offline batch inference.

This article reviews the top open-source small language models (SLMs) in 2026, including Qwen3.5-0.8B, Gemma-3n-E2B-IT, Phi-4-mini-instruct, SmolLM3-3B, and Ministral-3-3B-Instruct-2512. It discusses their suitability for production in resource-constrained environments, pros and cons, and answers common FAQs about SLMs.

This article explores the top open-source image generation models in 2026, including FLUX.2, Stable Diffusion, GLM-Image, and Z-Image-Turbo, highlighting their strengths, considerations, and use cases.

This article explains GPU memory (VRAM) in the context of LLM inference, covering how memory is used for model weights, KV cache, and overhead. It provides formulas for memory estimation, discusses common pitfalls like OOM errors, and presents optimization strategies such as quantization, distributed inference, and KV cache optimizations. The post also highlights how the BentoML Inference Platform simplifies these optimizations.

This guide explains the differences among DeepSeek-V3, R1, V3.1, and their variants, including performance benchmarks, use cases, and deployment tips.

This article covers the best open-source large language models in 2026, including DeepSeek-V4, MiMo-V2.5-Pro, and Kimi-K2.6, and answers common FAQs about performance, inference optimization, and self-hosted deployment.

This guide details ChatGPT usage limits as of April 2026 across Free, Go, Plus, Business, and Pro plans, explaining message caps, model selection, and context windows. It covers why limits exist (infrastructure load, cost control, fairness, abuse prevention) and other limitations like unpredictable performance, data privacy, lack of customization, and spiraling costs. The solution proposed is self-hosting open-source LLMs to remove all restrictions.

A comprehensive guide on training custom EAGLE-3 draft heads for speculative decoding to achieve 1.5-2.5x latency improvements in LLM inference without sacrificing output quality. Covers dataset preparation, hyperparameter tuning, training workflow, evaluation, and deployment.

Three universal patterns to optimize AI harnesses: keep .md files lean and human-written, use the R.P.I. (Research, Plan, Implement) framework for structured prompts, and employ subagents (parallel fan-out and pipelines) to maintain clean context. Emphasizes that the harness, not just the model, is where engineering judgment makes a difference, and advises committing to one harness and iterating rather than switching frequently.

NVIDIA Nemotron 3 Nano Omni is an open multimodal foundation model that unifies audio, images, video, and text into a single context. Built on the Nemotron 3 Nano backbone, it powers sub-agents in agentic workflows with leading efficiency and accuracy. Baseten now supports the model with day-zero availability, high-performance inference, multi-cloud capacity management, and enterprise security.

Baseten launches Frontier Gateway, a managed API gateway for AI labs to serve models under their own domain without building or buying a separate gateway, leveraging Baseten's inference infrastructure for performance and scalability.

DFlash introduces block diffusion for speculative decoding, predicting multiple tokens in parallel to surpass EAGLE's ~2x speedup ceiling. Baseten's implementation achieves ~3x speedups on Qwen3-8B across benchmarks, 10-30% faster than vLLM.

Baseten optimized Qwen3-TTS to achieve $3 per million characters, 90% lower cost than closed-source alternatives, with high fidelity for voice agents and voice cloning. The post details optimizations like disaggregated acoustic token generation/decoding, CUDA graphs, speaker embedding caching, dynamic frame accumulation, and word timestamps.