Models

A machine learning platform for Apple Silicon, written in Rust, covering local LLM inference, LoRA/QLoRA fine-tuning, quantization and serving through a desktop GUI, a terminal TUI and Metal/Neural Engine acceleration.

• Models
• Inference

compressed-tensors is a library that extends the safetensors checkpoint format to store and load tensors compressed with quantization and sparsity schemes such as GPTQ, AWQ, SmoothQuant and SparseGPT in one unified format.

• Models
• Libraries

A family of on-device embodied-AI models for robots: a 1.5B generalist manipulation policy and a 0.9B on-device target tracker, both designed to run locally without a cloud connection.

• Models

An MLX port of Moonshot AI's 2.78-trillion-parameter Kimi-K3 multimodal mixture-of-experts model for Apple Silicon, with a streaming weight converter, expert pruning and per-language expert-overlap analysis, though no single Mac can currently run it.

• Models
• Inference

An INT4 weight-quantization method for LLMs that uses learned pairwise rotations to suppress weight outliers, closing most of the accuracy gap to FP16 while running close to AWQ speed on NVIDIA GPUs and Apple Silicon.

• Models
• Papers

A research codebase for fine-tuning flow-matching robot control policies with online reinforcement learning after an initial imitation-learning stage, evaluated on locomotion and manipulation benchmarks.

• Models
• Papers

A generalized biological foundation model that unifies nucleic acid and protein language modeling in a single architecture, usable for sequence embedding, masked pretraining, and sequence- or token-level classification and regression tasks.

• Models

An open, compact language model series (4B and 1.7B parameters) tuned for coding, agentic tool use, and everyday tasks, with a native context window of up to one million tokens and broad hardware and framework support.

• Models

A research framework for in-context reinforcement learning that internalizes reusable skills into an agent's policy, evaluated on tasks such as ALFWorld and Search-QA.

• Models
• Papers

A reinforcement fine-tuning framework for large language and vision-language models, covering reinforcement learning with verifiable rewards, RLHF, and reward-model-driven policy optimization on top of PyTorch distributed training and rollout backends.

• Models

Hand-written PyTorch implementations of modern open-source LLM architectures such as Llama, Qwen, DeepSeek and Gemma, built one model at a time for clarity and learning rather than production use.

• Models
• Tutorials

A training toolkit for fine-tuning large language models locally on Apple Silicon with MLX, supporting LoRA, DoRA, full-precision and quantized training across many preference-optimization algorithms.

• Models

A family of open-source large language models tuned to better understand Southeast Asia's diverse languages and cultures, developed by AI Singapore's Products Pillar.

• Models

A NeurIPS 2026 paper and reference implementation showing that standard vision-language models can be trained to do object-level 3D understanding, depth estimation, pixel correspondence and camera pose estimation without special architectures or losses.

• Models
• Papers

A Python package that automatically compresses large language models, detecting available GPU memory, choosing a bit-width per layer, quantizing with error correction, and evaluating the result in one command.

• Models

A reinforcement-learning framework that trains language-model agents inside a central-agent multi-agent harness without flattening execution into a single token sequence, routing gradients to action and argument tokens separately.

• Agent frameworks
• Models

An open-source toolkit for quantizing and compressing PyTorch and Keras neural networks for efficient, constrained hardware, with post-training quantization, gradient-based fine-tuning and quantization-aware training.

• Models

A catalog of downloadable, verified model conversions and their conversion recipes for Apple's on-device Core AI runtime on iPhone and Mac, each carrying the recipe and validation evidence used to produce it.

• Models

A complete pipeline that trains a tiny decoder-only Transformer, from a randomly initialized tokenizer and model, on a person's own iMessage history, entirely on a Mac, to learn their phrasing and common replies.

• Models
• Tutorials

An LLM post-training framework for reinforcement-learning scaling, built on slime's training and data-generation design, that uses vLLM with a router as its default rollout backend.

• Models

A config-driven knowledge-distillation toolkit that turns a black-box teacher model into training data, composing generation, evaluation, filtering, rewriting, balancing and preference-scoring operators into pipelines that build SFT and DPO datasets.

• Models

A dual-architecture research framework that unifies the exact output of autoregressive large language models with the parallel, high-speed token generation of diffusion models, for strictly lossless faster inference.

• Models
• Inference

Robust-U1 is the official implementation of a unified multimodal language model that detects and repairs corrupted visual input before reasoning about it, aimed at robust visual understanding under real-world image degradation.

• Models

A one-click, portable Windows installer for AI-Toolkit by Ostris, a training suite for image and video diffusion models, setting up an embedded Python, Git, Node.js, FFmpeg, PyTorch and a pinned Triton build with no system Python needed.

• Models
• Generative media