Poolside names Laguna XS.2 hackathon winners
Poolside AI hosted a weekend model research hackathon in London in partnership with NVIDIA, Prime Intellect, and Hugging Face, bringing together around 30 teams to build on top of Laguna XS.2, their recently launched 33B Mixture-of-Experts (MoE) agentic coding model. The top prize went to Emil Fristed of Overthinking Machines Labs for a pseudo-full-duplex dialogue reasoning method using silence tokens. Second place was awarded to 'Coding Kernels by the Pool' for compiling PyTorch to CUDA using a dense distillation of Laguna, while third place was taken by Alara Dirik for KV cache product vector quantization, and an honorary mention was given to Aaron Kazah for adding SigLIP-based vision capabilities.
While open-weight models are typically used out-of-the-box, Poolside's hackathon demonstrates that Laguna XS.2's highly efficient MoE structure makes it a uniquely powerful and accessible playground for advanced post-training, reinforcement learning, and architectural experimentation.
* By activating only 3B parameters out of 33B total, Laguna XS.2 offers developers a high-reasoning model that runs locally on commodity hardware while maintaining competitive software-engineering benchmarks.
* The winning hacks showcased impressive post-training diversity, ranging from pseudo-full-duplex dialogue and PyTorch-to-CUDA compilation, to KV cache quantization and visual model adapters.
* Collaborations between model builders like Poolside and compute/infra providers like NVIDIA and Prime Intellect demonstrate how rapidly reinforcement learning and fine-tuning are democratizing for independent researcher groups.
DISCOVERED
1h ago
2026-06-01
PUBLISHED
1h ago
2026-06-01
RELEVANCE
AUTHOR
PrimeIntellect
