In this tutorial, we explore how to build a Context-Folding LLM Agent that efficiently solves long, complex tasks by intelligently managing limited context. We design the agent to break down a large task into smaller subtasks, perform reasoning or calculations when needed, and then fold each completed sub-trajectory into concise summaries. By doing this, we…
What would you build if you could run Reinforcement Learning (RL) post-training on a 32B LLM in 4-bit NVFP4—on a single H100—with BF16-level accuracy and 1.2–1.5× step speedups? NVIDIA researchers (with collaborators from MIT, HKU, and Tsinghua) have open-sourced QeRL (Quantization-enhanced Reinforcement Learning), a training framework that pushes Reinforcement Learning (RL) post-training into 4-bit FP4…
In this tutorial, we explore Ivy’s remarkable ability to unify machine learning development across frameworks. We begin by writing a fully framework-agnostic neural network that runs seamlessly on NumPy, PyTorch, TensorFlow, and JAX. We then dive into code transpilation, unified APIs, and advanced features like Ivy Containers and graph tracing, all designed to make deep…
Anthropic released Claude Haiku 4.5, a latency-optimized “small” model that delivers similar levels of coding performance to Claude Sonnet 4 while running more than twice as fast at one-third the cost. The model is immediately available via Anthropic’s API and in partner catalogs on Amazon Bedrock and Google Cloud Vertex AI. Pricing is $1/MTok input…
ServiceNow Research has released DRBench, a benchmark and runnable environment to evaluate “deep research” agents on open-ended enterprise tasks that require synthesizing facts from both public web and private organizational data into properly cited reports. Unlike web-only testbeds, DRBench stages heterogeneous, enterprise-style workflows—files, emails, chat logs, and cloud storage—so agents must retrieve, filter, and attribute…
Tuning LLM outputs is largely a decoding problem: you shape the model’s next-token distribution with a handful of sampling controls—max tokens (caps response length under the model’s context limit), temperature (logit scaling for more/less randomness), top-p/nucleus and top-k (truncate the candidate set by probability mass or rank), frequency and presence penalties (discourage repetition or encourage…
NVIDIA AI has introduced Reinforcement Learning Pretraining (RLP), a training objective that injects reinforcement learning into the pretraining stage rather than deferring it to post-training. The core idea is simple and testable: treat a short chain-of-thought (CoT) as an action sampled before next-token prediction and reward it by the information gain it provides on the…
In this tutorial, we explore the advanced capabilities of PyTest, one of the most powerful testing frameworks in Python. We build a complete mini-project from scratch that demonstrates fixtures, markers, plugins, parameterization, and custom configuration. We focus on showing how PyTest can evolve from a simple test runner into a robust, extensible system for real-world…
Andrej Karpathy has open-sourced nanochat, a compact, dependency-light codebase that implements a full ChatGPT-style stack—from tokenizer training to web UI inference—aimed at reproducible, hackable LLM training on a single multi-GPU node. The repo provides a single-script “speedrun” that executes the full loop: tokenization, base pretraining, mid-training on chat/multiple-choice/tool-use data, Supervised Finetuning (SFT), optional RL on…
Do you actually need a giant VLM when dense Qwen3-VL 4B/8B (Instruct/Thinking) with FP8 runs in low VRAM yet retains 256K→1M context and the full capability surface? Alibaba’s Qwen team has expanded its multimodal lineup with dense Qwen3-VL models at 4B and 8B scales, each shipping in two task profiles—Instruct and Thinking—plus FP8-quantized checkpoints for…
