Open Source C AI Models

The llama.cpp project enables the inference of Meta's LLaMA model (and other models) in pure C/C++ without requiring a Python runtime. It is designed for efficient and fast model execution, offering easy integration for applications needing LLM-based capabilities. The repository focuses on providing a highly optimized and portable implementation for running large language models directly within C/C++ environments.

Run a fast ChatGPT-like model locally on your device. This combines the LLaMA foundation model with an open reproduction of Stanford Alpaca a fine-tuning of the base model to obey instructions (akin to the RLHF used to train ChatGPT) and a set of modifications to llama.cpp to add a chat interface. Download the zip file corresponding to your operating system from the latest release. The weights are based on the published fine-tunes from alpaca-lora, converted back into a PyTorch checkpoint with a modified script and then quantized with llama.cpp the regular way.

FLUX.2-klein-4B is a compact, high-performance C library implementation of the Flux optimization algorithm — an iterative approach for solving large-scale optimization problems common in scientific computing, machine learning, and numerical simulation. Written with a strong emphasis on simplicity, correctness, and performance, it abstracts the core logic of flux-based optimization into a minimal C API that can be embedded in broader applications without pulling in heavy dependencies. Because the implementation is in plain C and focuses on data locality and vectorized operations, flux2.c can be integrated into performance-critical code paths where control over memory layout and execution behavior matters, such as GPU kernels, embedded systems, or custom ML runtime engines.

Colibri is a compact inference engine designed to run the 744-billion-parameter GLM-5.2 mixture-of-experts model on consumer hardware. It keeps the dense portion of the quantized model in memory while streaming routed experts from a large disk-based store as they are needed. The runtime is implemented in pure C, requires no Python or BLAS during inference, and can operate without a GPU. Compressed attention caches, expert caching, optional hot tiers, and speculative decoding reduce memory pressure and improve repeated use. A planning tool calculates safe disk, RAM, and VRAM placement before loading the model, while a diagnostic command checks system readiness. Colibri includes terminal chat, an OpenAI-compatible text API, and a browser client, but disk-bound generation can be slow on cold caches.

Kimi K3 in C is a portable C99 inference engine built to run the 2.78-trillion-parameter Kimi K3 model on CPUs without BLAS, machine-learning frameworks, or GPUs. It demonstrates inference from a roughly 1.56 TB checkpoint with measured memory use as low as 8.24 GB. The runtime streams model trunk layers and routed experts from disk instead of keeping all weights resident in memory. Memory presets balance pinned layers and an expert LRU cache for laptops, desktops, workstations, and servers. Incremental generation preserves KV cache and recurrent state between tokens. The repository also includes tokenization, safetensors loading, diagnostics, benchmarks, trace replay, and comparisons against a PyTorch reference.

Seamless Communication is a research project focused on building more integrated, low-latency multimodal communication between humans and AI agents. The motivation is to move beyond “text in, text out” and enable direct, live, multi-turn exchange involving language, gesture, gaze, vision, and modality switching without user friction. The system architecture includes a real-time multimodal signal pipeline for audio, video, and sensor data, a dialog manager that can decide when to act (speak, gesture, point) or query, and a cross-modal reasoning layer that fuses perception with semantic context. The research prototype includes components for visual grounding (understanding when a user references something in view), gesture recognition and synthesis, and turn-taking mechanisms that mirror human conversational timing. Because latency and synchronization are critical, the codebase invests in asynchronous scheduling, overlap of perception and reasoning, and fast fallback responses.

WASTE is an embeddable C inference engine for running extremely large mixture-of-experts models when the weights exceed available RAM. It keeps the shared model trunk in memory and streams only the experts selected for each token from fast NVMe storage. A bounded cache reuses recently needed experts, while lookahead routing begins disk reads before the next layer requires them. Its main target is the full 2.78-trillion-parameter Kimi K3 model, including multimodal image input. The engine has no third-party runtime dependency on its CPU inference path and exposes both a CLI and C library. An optional server provides an OpenAI-compatible chat API with streaming, tools, structured output, and image support. Validation compares layers, logits, vision output, and prompt rendering against reference implementations.

fairseq2 is a modern, modular sequence modeling framework developed by Meta AI Research as a complete redesign of the original fairseq library. Built from the ground up for scalability, composability, and research flexibility, fairseq2 supports a broad range of language, speech, and multimodal content generation tasks, including instruction fine-tuning, reinforcement learning from human feedback (RLHF), and large-scale multilingual modeling. Unlike the original fairseq—which evolved into a large, monolithic codebase—fairseq2 introduces a clean, plugin-oriented architecture designed for long-term maintainability and rapid experimentation. It supports multi-GPU and multi-node distributed training using DDP, FSDP, and tensor parallelism, capable of scaling up to 70B+ parameter models. The framework integrates seamlessly with PyTorch 2.x features such as torch.compile, Fully Sharded Data Parallel (FSDP), and modern configuration management.