Models

A from-scratch large language model stack, covering the model, an instruction set, a compiler, a simulator, and a GPU design, built small enough for one person to read end to end.

• Models
• Tutorials

The official code and project write-up for an ICML 2026 paper showing that smaller language models compress their token embeddings into a narrow region, and proposing a dispersion loss to counteract it.

• Models
• Papers

A LoRA fine-tune of Gemma that answers in a terse, compressed style to cut output tokens, while keeping code blocks and meaning intact, evaluated on held-out prompt pairs.

• Models

A 25-language speech recognition model built from NVIDIA's Parakeet TDT 0.6B v3, replacing half its encoder's temporal filters with fitted, frozen Gabor kernels and retraining the rest, reporting lower error rates than Parakeet on most benchmarks.

• Models
• Speech

A diffusion language model that decodes many tokens in parallel by revising its own earlier predictions, aiming to preserve output quality at high decoding parallelism.

• Models
• Inference

A Jax library for quantization-aware training and post-training quantization of Flax models, targeting CPUs, GPUs and TPUs as well as on-device LiteRT deployment.

• Models

A cross-platform quantization engine for PyTorch models that reduces memory use and increases inference and training speed on GPUs, NPUs, CPUs and ARM devices.

• Models
• Inference

A 2-bit complex-valued large language model in which every weight is one of four complex values, achieving efficient compression with minimal accuracy loss compared with real-valued 2-bit models.

• Models

An open-source omnimodal large language model that combines image, speech and text understanding with real-time, emotionally expressive speech generation.

• Models
• Speech

An RL post-training framework for retrieval-augmented reasoning that adds explicit knowledge-refinement steps between search calls, from a NeurIPS 2025 paper.

• RAG
• Models

An alpha-stage TypeScript framework for fine-tuning open-weight LLMs with type-safe configs, a local Studio web UI, and managed GPUs.

• Models

A Python implementation of the PALI-3 vision-language model, which encodes images with a contrastively pretrained SigLIP vision transformer and generates text through a UL2 encoder-decoder transformer.

• Models

A frozen-task, frozen-hardware public leaderboard for speedrunning LoRA fine-tuning of small language models, with every record independently re-verified on a cloud GPU sandbox.

• Models
• AI evals

A curated collection of datasets for training and evaluating large language models, covering medical AI, NLP, multimodal learning, instruction tuning, reasoning, code generation and evaluation benchmarks.

• Models
• Curated lists

An open-source Python framework for fine-tuning retrieval-augmented generation systems across both centralized and federated architectures.

• RAG
• Models

A reference implementation of a ternary-weight tiny language model with a bit-exact Python-to-C99 inference engine, sized to run inside a microcontroller's RAM budget.

• Models
• Inference

A reinforcement-learning post-training framework that keeps trillion-parameter language models at low-precision deployment format and trains only a small BF16 adapter, fitting frontier-scale RL onto a single node.

• Models

A family of methods that compress or actively manage a language model's long reasoning traces into compact representations, cutting the cost of retaining and attending to them during inference.

• Models

An unofficial PyTorch reproduction of DeepSeek's method for teaching multimodal LLMs to reason with bounding boxes and points interleaved directly in their chain of thought.

• Models

A command-line tool implementing Retrieval-Augmented Fine-Tuning, fetching a person's public writing from sources such as Substack, X, or RSS to build a combined fine-tuning and retrieval dataset.

• RAG
• Models

A Terraform module that deploys an end-to-end pipeline for extracting question-and-answer pairs from documents and fine-tuning a large language model on them using Google Cloud Vertex AI.

• Models
• Hosting

A Swift library for local computer-use on Apple Silicon: it reads a form in a running app or browser through the Accessibility API and types answers into each field using a small on-device model.

• Models
• Libraries

An extension to ONNX with custom operators for arbitrary-precision integer and minifloat quantization, plus Python utilities for executing, transforming and analyzing quantized neural network models.

• Models
• Libraries

The official research codebase for a paper showing that predicting the next latent state, instead of the next token, trains transformers to learn more compact world models.

• Models
• Papers