Leanstral: Open-Source foundation for trustworthy vibe-coding
Get smart on it
Leanstral is an open-source AI code agent designed specifically to work with Lean 4, a proof assistant used for expressing mathematical objects and software specifications. The tool addresses a scaling bottleneck in using AI for high-stakes coding by enabling agents to formally prove their implementations against strict specifications rather than requiring humans to manually debug machine-generated code. Leanstral uses a sparse architecture with 6 billion active parameters and is trained to work efficiently in realistic formal repositories, while supporting integration with other tools through MCPs. In benchmarks using a new evaluation suite called FLTEval, Leanstral demonstrated cost efficiency advantages over both larger open-source models and closed-source competitors, achieving competitive or superior scores at a fraction of the computational cost.
Related news
Vibe coding platform Base44 launches own model as AI startups seek defensibility
Wix-owned vibe coding platform Base44 has started rolling out its own AI model — with hopes that it will eventually outperform frontier models.
DiScoFormer: One transformer for density and score, across distributions
DiScoFormer is a transformer model that estimates both the density and score of a data distribution in a single forward pass without requiring retraining for new distributions. The density describes where data points cluster, while the score, the gradient of log-density, points toward more probable regions and is used in diffusion-based generative models, Bayesian sampling, and scientific simulations. Existing methods force a trade-off: classical kernel density estimation works on any distribution but loses accuracy in high dimensions, while neural score-matching models stay accurate in high dimensions but must be retrained for each new distribution. DiScoFormer significantly outperforms kernel density estimation, cutting score error by 6.5 times and density error by more than 37 times in 100 dimensions, while generalizing to distributions with shapes and complexities not seen during training.
