Topic
All digests tagged Model Optimization
Suggested route
Move from earlier context through the available material to the newest signals. This route is a reading aid, not a difficulty ranking.
A faster model only helps if the finished job gets faster. We test M3.1 Flash by measuring the total cost (time, compute, and manual fixes) of a complex research report, proving that speed must be balanced with accuracy and reliability.
Read article
This talk covers the systems and algorithmic challenges behind making globally distributed RL work at scale. We will discuss how we decouple training from inference, propagate compressed weight updates across datacenters, and design the infrastructure to tolerate frequent hardware failures.
Read article
Running open LLMs in production requires managing the full loop: inference, data, post-training, and deployment. Nebius Token Factory provides a full-stack solution, optimizing performance using techniques like speculative decoding and KV cache offloading.
Read article
Learn how autoresearch, custom GPU kernels, and bare-metal hacks can achieve 3x model inference speedups. Discover the critical role of human ideas versus agent tuning.
Read article
Learn how Per-Layer Embeddings (PLE) enhance Gemma 4 by giving each token a unique, layer-specific embedding. This boosts model performance and diversity without increasing computational parameters.
Read article
NVIDIA experts detail the Nemotron post-training pipeline, explaining how structured data, NeMo Gym, and MOPD are used to build agentic LLMs. Learn about data bias mitigation and the evolution from pre-training.
Read article
TensorRT Model Connect simplifies the path from open-source models to deployable applications. It provides a consistent workflow to convert models into optimized bundles, supporting everything from LLMs to video generation on multi-device setups.
Read article
This paper proposes Direct On-Policy Distillation, a cheap way to transfer RL benefits from small models to large ones by using policy shifts as dense reward signals.
Read article
Open models are key to building trust and sovereignty. By owning the full stack—from weights to training traces—developers can specialize AI for niche use cases, achieving performance that surpasses generalized frontier closed APIs.
Read article
The research team discusses a joint scaling law for pre-training and RL, showing how compute allocation across these stages impacts downstream task performance. Key findings relate to optimizing model size versus training duration.
Read article
Dive into train-time scaling and RL techniques that enable self-improving AI agents. Learn how STaR, DeepSeekMath, and DAPO boost reasoning in LLMs using closed feedback loops.
Read article