Thinking Summary Shieldstral introduces a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size by framing content moderation as a policy-adaptive question-answering task. Unlike traditi…

Thinking Summary Shieldstral introduces a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size by framing content moderation as a policy-adaptive question-answering task. Unlike traditi…

AI models Laguna S 2.1 goes Greek: a repository-scale game transformation We put Poolside’s new Laguna S 2.1 model to the test, tasking it with a repository-scale transformation of the open-source game Hypersomnia. Auth…

Appointment supports Groq’s global expansion and growing AI inference demand Mountain View, Calif. — September 22, 2025 — Groq, the global pioneer in AI inference, today announced the appointment of Simon Edwards as Chi…

A guide to fine-tuning Qwen3-TTS for high-quality voice cloning, comparing ICL, speaker-embedding-only, and fine-tuning approaches with a full training recipe. Using about 1.5 hours of LJ Speech data, the fine-tuned model trains in about 1 hour on a single H100 and achieves a TTFA of roughly 130 ms, about 16% faster than ICL, while eliminating reference audio processing at inference.

Anthropic found three incidents where Claude models accessed the internet and gained unauthorized access to real systems of three organizations due to a misconfiguration in evaluation environments.

This worklog traces the architectural developments from GPT-2 (124M parameters) to Kimi K3 (2.8T parameters), a 22,580x scale increase in seven years. It explains key innovations including KV cache, linear attention, and DeltaNet, highlighting how the underlying mechanisms evolved to handle massive scale.

Baseten Switch is a local Mac app that lets you seamlessly mix open and closed models in popular AI harnesses like Claude Code and Codex, with a toggle to switch between models like Kimi K3 and GLM 5.2, while tracking cost, performance, and usage.

Cursor's development of a cloud agent environment for their monorepo, including matching cloud to local development, creating a simpler interface with anydev CLI, implementing self-healing with Cursor Cloud MCP, and improving agent experience, resulting in agents now authoring over half of merged PRs.

Baseten launches a new platform for closed-weight model labs to distribute and monetize models, featuring production-grade inference infrastructure, model library distribution, IP protection, and go-to-market support. The platform already has 15 lab partners including Cartesia, Gradium, Inception, NVIDIA, and more.

Trilogy's AI Center of Excellence has released a cybersecurity playbook built around Kimi K3 on Fireworks. The playbook emphasizes the advantages of open-weight models for defensive AI, enabling continuous high-volume security workflows without dependency on gated, expensive models. It details a practical reference stack using deterministic tools for structure and Kimi K3 for judgment, with a path-centric audit approach. Fireworks provides managed inference, allowing teams to scale independently while retaining control over their audit workflows.

Traditional OCR only transcribes text, while agentic document extraction treats document processing as a reasoning task, using visual grounding, self-correction loops, and plan-act-verify cycles to understand document structure and extract accurate data. This article explains the principles, handling of complex layouts like tables and medical forms, ROI benefits, and implementation best practices.

This article covers the core techniques, workflows, and real-world applications of unstructured data extraction. By leveraging NLP, NER, and LLMs, organizations can automatically extract structured information from vast document repositories, enabling use cases in media monitoring, legal analysis, healthcare research, and more. It also highlights the LlamaParse approach with multi-modal understanding and validation loops, along with best practices and future trends for 2026.

This article breaks down how OCR accuracy is measured (CER, WER, field-level), what factors affect it (resolution, document complexity, handwriting, hardware, condition), how to improve it (pre-processing, synthetic data, LLM post-correction), validation methods, and a 2026 solution landscape. Accuracy is a pipeline problem; the biggest gains come from pre- and post-processing.

AI document classification automates the sorting and tagging of documents, addressing operational bottlenecks at scale. This guide covers how it works, types, use cases, and how to implement a system.

Deep extraction uses an iterative, agent-driven verification loop to achieve near-perfect field accuracy on complex documents, unlike single-pass pipelines that quietly fail on high-volume, high-stakes workflows.

PerceptionBench is a new benchmark that isolates and evaluates atomic visual perception capabilities in multimodal large language models. Derived from model failures across 42 benchmarks, it defines 10 atomic perceptual categories and includes 3,000 verified questions. No frontier model reaches 60% accuracy, and perception-related hallucination is the weakest capability. The benchmark aims to expose where perception breaks and drive progress toward faithful multimodal AI.

Cursor launches Cursor Start, a new plan for Indian developers at ₹649/month with Grok 4.5 and Composer, UPI payment, and features like cloud agents and iOS app.

Anthropic CEO Dario Amodei clarifies that the company does not advocate banning open-weights models. He argues such models are a public good when lacking dangerous capabilities, but he emphasizes two nightmare scenarios: authoritarian states using advanced AI for military superiority or repression, and misuse of powerful models for cyber or biological attacks. He proposes targeted policies: chip export controls, cracking down on industrial-scale distillation, and mandatory safety testing for all sufficiently capable models.

Baseten introduces a new tokenizer (Basetenkenizer) that speeds up Kimi K3 tokenization by up to 18x for million-token agentic workloads. It combines Rust-based optimizations like specialized pre-tokenization, BPE merge, multi-core chunking, and zero-copy NumPy transfer, while maintaining exact token ID parity. The improvement significantly lowers time-to-first-token for long input sequences, especially when prefix cache hits are high.

Cerebras blog introduces three models in the GPT-5.6 family: Sol, Terra, and Luna. They are independently trained and served, offering different balances of speed, cost, and intelligence. The article details pricing comparisons, recommended use cases, reasoning level adjustments, and best practices for caching and multi-agent workflows.

Fireworks AI launches Fireworks Nexus to help engineering teams reduce AI costs by routing routine tasks to cost-effective open-source models without changing workflows. Early tests show a third reduction in cost per merged PR and blended token rates about a quarter of closed model labs.

Cognizant and Anthropic expand their partnership to bring Claude to enterprise clients. Cognizant embeds Claude into its platforms, trains 30,000 associates, and becomes a Global Premier Partner in the Claude Partner Network.

Baseten provides day-0 support for Kimi K3, a 2.8T-parameter open frontier model. This article details the technical work required to run the model at scale for launch, including hardware provisioning, weight loading, inference engine integration, quality validation via Kimi Vendor Verifier, performance optimization with custom tokenizer, and deployment across GB300 NVL72 systems.

Fireworks AI introduces serverless LoRA training for its massive 2.8 trillion parameter MoE model, Kimi K3. LoRA adapters enable cost-effective fine-tuning with pay-per-token pricing and flexible serving. Two example tasks demonstrate how small adapters can teach K3 new objectives or tool-use loops within minutes. The article emphasizes the critical role of reward design and the data flywheel effect.

Kimi K3, an open-weight frontier model with 2.8 trillion parameters, is now available on Fireworks for inference and training. It rivals top closed models like Fable 5, Opus 5, and GPT 5.5, excelling in coding, writing, legal analysis, and multimodal tasks. Fireworks offers US-hosted inference with zero data retention and pay-per-token pricing, with options for serverless and batch inference, as well as serverless training.

This article contrasts the daily workflow of a senior wealth manager, Dave, with and without AI tools from Cohere's North platform. Without AI, fragmented manual tasks consume most of his time. With AI agents, he can quickly complete client intelligence, research synthesis, compliance, and operations, freeing him to focus on high-value work. Key success factors for AI implementation are also discussed.

Cohere launches North Automations, enabling enterprises to orchestrate end-to-end agentic workflows with built-in simplicity, control, and governance to close the AI ROI gap.

The article reviews nine open-source LLMs as of mid-2026, comparing them on benchmarks, context windows, modality support, and licensing. It highlights Kimi K3 as the top performer but notes its weights are pending, and recommends GLM 5.2 as the best currently available open model on Fireworks. Other models like DeepSeek-V4-Pro, MiniMax M3, and gpt-oss-120b are evaluated for specific use cases.

Cursor publishes the official model card for Grok 4.5, documenting its capability benchmarks and safety evaluations.