Cognisoc builds open-source runtimes and tooling that run large language models on your own hardware — from a bare-metal unikernel to a phone in your pocket, with native bindings across eight languages. No cloud required.
What is Cognisoc?
Cognisoc is an open-source effort to make running large language models locally as easy as calling a cloud API — in whatever language, on whatever hardware you already have. Rather than a single tool, it is a family of purpose-built runtimes, each solving one layer of the inference problem: a polyglot server, mobile and .NET runtimes, a composable Rust engine, an educational implementation, and a bare-metal unikernel. Everything is open source and built on the GGUF / llama.cpp ecosystem, so models are portable across the whole stack.
Six open-source tools, each with its own home on the web. Pick the layer you need — they all speak GGUF.
Runtime + server
Local LLM runtime in Rust, built on llama.cpp, with native bindings for Rust, Python, Node.js, Go, PHP, and C/C++ — and drop-in Ollama CLI compatibility.
Who it's for: Polyglot teams and app developers who want one inference dependency across many languages.
Mobile
A Flutter FFI plugin, built on llama.cpp, for on-device GGUF model inference on Android and iOS — no cloud, no latency, complete privacy.
Who it's for: Mobile developers shipping private, offline AI features in Flutter apps.
Runtime / engine
A modular LLM inference runtime in Rust with a unified, type-safe interface across 47 architectures, built on three composable cores.
Who it's for: Rust engineers who want a composable, type-safe inference engine to build on.
Runtime (.NET)
A .NET-native runtime for local GGUF inference — managed-by-default, with optional Metal and Vulkan compute backends.
Who it's for: C# and .NET developers who want local inference with one package and no native setup.
Education
Learn how LLMs work by building one in Zig — an educational, book-shaped codebase implementing 18 transformer families across 6 progressive layers.
Who it's for: ML/systems-curious learners and Zig fans who learn by reading and writing code.
Bare metal
A Multiboot unikernel in C that boots on bare metal or in QEMU and serves an llama.cpp-compatible HTTP API — the kernel is the application.
Who it's for: Systems engineers exploring OS-free, boots-to-inference serving on bare metal.
The capabilities that run across the portfolio — the reasons to reach for a Cognisoc tool.
First-party native APIs for Rust, Python, Node.js, Go, PHP, C/C++, Dart, and C#/.NET — not just an HTTP wrapper.
GGUF inference that runs on servers, desktops, and mobile — no cloud round-trip and no data leaving the device.
Drop-in Ollama CLI, plus OpenAI- and Anthropic-compatible HTTP APIs, so existing tooling keeps working.
A Multiboot unikernel that boots straight to an HTTP inference API with no operating system underneath.
A readable, book-shaped Zig implementation that teaches transformers from tensors to text, layer by layer.
Developed in the open on GitHub under permissive licenses (mostly MIT/Apache-2.0). Infrastructure as a public good.
Start from what you're trying to do — each row links out to the right product.
| I want to… | Reach for | |
|---|---|---|
| Replace Ollama but keep native bindings in 6 languages | mullama | Learn more → |
| Run a model entirely on a phone (iOS / Android) | llamafu | Learn more → |
| Embed inference in a C# / .NET app with no native setup | llmdot | Learn more → |
| Build on a composable, type-safe Rust inference engine | unillm | Learn more → |
| Actually understand how transformers work, in code | zigllm | Learn more → |
| Serve inference from bare metal with no host OS | cllm | Learn more → |
The principles behind every tool in the portfolio.
Inference should run where your data already is — on your servers, desktops, and phones — not behind a metered cloud API.
Every project is developed in the open under permissive licenses. The inference layer is a public good, not a proprietary moat.
Native bindings for eight languages and ecosystems, so teams reach for one runtime instead of gluing HTTP calls together.
From a bare-metal unikernel to a Flutter phone plugin, we cover every layer of the inference stack with purpose-built tools.
Explore the products, read the developer guides, or reach out about research and hardware collaboration. Everything is open source on GitHub.