A popular machine learning framework for tensors and deep learning applications.

A universal document converter for converting between various markup formats.

Allows running 70B LLMs on single consumer GPUs with as little as 4GB VRAM.

A framework for running ML models inside DRAM by breaking DDR4 timing rules.

Largest open-access wearable health dataset and open-source wearable foundation models

A CLI framework for deterministic replay of AI agent executions using a MITM proxy for isolation.

The largest open-weight LLM to date with 2.8 trillion parameters.

An AI-native, multimodal workbench for scientific discovery.

A 2.8-trillion-parameter open-source frontier-class LLM with a 1-million-token context window.

An authorization gateway for AI agents that uses identity-bound signatures to prevent unauthorized tool calls.

A lightweight, high-quality text-to-speech model for local execution.

An open-source, all-atom biomolecular foundation model for drug discovery.

A system for serving map tiles from a single static file (pmtiles).

Codebase for training and evaluating speculative decoding systems, supporting DSpark.

A tool for running open-source LLMs locally

Cloud service for Lean 4 proof manipulation, extraction, and verification.

A port of Llama models to C/C++ for efficient on-device inference.

C++/CUDA training framework featuring native autograd, caching allocator, and MLIR-based compiler.

Open-source MoE language models with 1M context support and high efficiency.

A controlled-natural-language prover for LLM-generated mathematics that outputs Lean files.