Poolside Research Hackathon
Always wanted to push the frontier of RL with a locally deployable AI model?
Come join us!
Poolside is hosting a two-day, in-person model research hackathon in London centered on Laguna XS.2 33B-A3B: our first open-weight model built for agentic coding and long-horizon software work which can run on an NVIDIA DGX SparkTM.
This is a research hackathon for people who want to work on the model itself, not build apps on top of it.
The track is Foundations: fine-tuning, post-training, computer use, multi-agent, custom RL environments, evaluation design, quantization, and inference optimization.
The winning submission will receive an NVIDIA DGX SparkTM to continue the work on-device after the event!
Who Should Apply
We’re bringing together a curated group of top researchers and technical builders with hands-on experience working with models. You don’t need to have worked with Laguna XS.2 before.
Relevant backgrounds include:
Fine-tuning or post-training models
RL environments, reward design, or evaluation design
Inference optimization, quantization, or serving
Agentic coding workflows, software engineering agents, computer use, or tool-using agents
Benchmarking, model evaluation, or applied ML experimentation
Multi-agent workflows and research
Challenge
Submissions should meaningfully improve Laguna XS.2 for an economically valuable task or a research workflow.
Here are some examples of what you could pursue with Laguna XS.2:
Fine-tuning, post-training, custom RL environments
Computer use, multi-agent
New capabilities — e.g. specialized code-completion or deep research
Novel fine-tuned variants on Prime Intellect's platform
Reusable RL environments or reward signals
Extended use cases like agentic coding clients or claw agents
Quantization, and inference optimization.
Aggressively quantized variants — 2-bit or even 1-bit, on-policy distillation
Inference improvements — adapters, distillation, serving
Partners
The hackathon arc runs end-to-end on NVIDIA: you experiment with Laguna XS.2 on NVIDIA infrastructure via Prime Intellect during the event, then take the DGX Spark home to keep prototyping, fine-tuning, and running inference long after the weekend wraps. The prize is designed to extend the engagement, not end it.
Training during the hackathon will run through Prime Intellect Lab: a hosted platform for agentic model improvement, powered by NVIDIA infrastructure and designed to abstract away the complexity of managing GPU clusters and training infrastructure.
We’re also partnering with Hugging Face as the open-source home for the event. A dedicated HF org will host all submissions so the work produced over the weekend stays available for the wider community to inspect, run, and build on. Hugging Face is also providing $100 HF credits for participants to use during or after the event.
Judging Criteria
Projects will be judged on whether they meaningfully improve Laguna XS.2 — for an economically valuable task or a novel research idea — either by:
Expanding model use cases (computer use, multi-agent coordination, evaluation design), or
Reducing cost and latency (optimizations, speed, quantization)
Judges will prioritise generalisability, reproducibility, and the strength of the technical contribution.
The panel will include Poolside researchers and engineers, alongside one external guest judge from the wider AI research and infrastructure ecosystem.
Prizes
The winning team takes home an NVIDIA DGX Spark, a personal AI supercomputer purpose-built for agentic AI development and deployment, with enough memory and compute to run, fine-tune, and serve models the size of Laguna XS.2 directly from your desk.
Second and third place teams will receive credits (for Prime Intellect and Hugging Face), and every participant will leave with Poolside swag.
Apply
The event runs from 3PM to midnight on Friday and 10AM to 5PM on Saturday.
Capacity is limited and applications are reviewed on a rolling basis.
You can apply as an individual or as a team of up to 4 people. If you apply individually, you'll have the option to meet teammates in our Discord channel before the event or in person on Friday.
Exact venue details, agenda, technical setup instructions, and Discord access will be shared with selected participants ahead of the event.