Today we're launching Slingshots // FOUR: Fourteen projects spanning open models and recipes, self-improving systems, new ways to accelerate research and strengthen decision-making, and more.
It's a fine time to be doing open source computer science research, and Batch 4 shows us why. Marin is redefining openness itself – publishing their weights, but also data, logs, and decisions that the team has made throughout their hero run. Atlantic-CC unveils the pretraining data curation strategy behind the run, built on a spec that increases coverage by discarding less data. And Contrastive Language Models (CLMs) emerge as an open contender to the latest decision models, matching Jev's computer-use, gaming, and tool-calling tasks at up to 9x speeds. Today, anyone can study how a frontier model gets made.
Researchers are also exploring new paths to self-improving systems, focusing on training, context management, and harness development that enable models to keep learning after they ship. Pedagogical RL works on training itself, SkyDiscover on the search over algorithms and systems, ACE on the context, and Exoharness and dotagents on the harness. Our very own microharness Headlong, a collaboration with MIT, is designed with persistent agency, and ideates on self-improvement even when nobody is around to prompt it.
Core to this batch is the belief that no piece of the stack should go underexplored. This round includes a fast, simple approach to writing kernels, serving and query infrastructure for multimodal models, scalable environments for training better predictive models, and a benchmark and retrieval system to power stronger scientific research.
We're proud to back this group of researchers advancing the frontier. Our support looks different for each project — whether that be compute, product strategy, or a place to get together and build. The result is good ideas, shipped in the open, that anyone can experiment with and take further.
Meet Slingshots // FOUR.
An open lab for frontier model development, where the code, data, experiments, and training logs behind every model are public as the work happens. Stanford University
A reinforcement learning framework that teaches AI models to generate plausible, successful reasoning trajectories they can learn from for challenging tasks. MITLaude Institute
A universal serving system designed to be modular and extensible while achieving state-of-the-art performance on any model architecture and across all modalities. Stanford UniversityCarnegie Mellon UniversityUniversity of Washington
A fully recursive agent and harness architecture that can safely edit every aspect of itself at runtime to get better at your tasks. UC Berkeley
A CUDA framework that simplifies writing fast, high-utilization deep learning kernels with tile-based primitives. Stanford University
A system one model for fast and generalizable decision-making. Stanford University
An open, self-improving framework that enables AI to autonomously build and optimize high-performance, trustworthy software systems. UC Berkeley
A benchmark for finding prior work that could inspire scientific research, and a retrieval system trained for this task. Stanford University
A query engine for executing semantic and agentic queries over enterprise-scale data lakes. MITUniversity of Wisconsin
A framework that lets LLM agents self-improve by evolving their contexts into detailed playbooks, without any weight updates. Stanford University
Scalable environments to evaluate and train predictive intelligence that helps people make more informed decisions. ELLIS Institute TübingenMax Planck Institute for Intelligent Systems
A 17T token model-based processing of the Common Crawl for training bigger and better open-source large language models. Stanford UniversityOpen Athena
A microharness for persistent agents. Laude InstituteMIT
Researchers: Tell us what you're working on and what you need to get it into the world. We accept projects year-round, deploy resources immediately, and announce new batches every few months. Apply here.