mullama
/features

everything in the runtime

One Rust core on top of llama.cpp, exposed six ways, wire-compatible with the tools you already use, and deployable from a laptop to a GPU box.

// bindings

One core, six idiomatic APIs

Rust

The native core. A safe Rust API over llama.cpp — Model, Context, ContextParams.

Python

pip install mullama. Idiomatic Python, not transliterated Rust; supports n_gpu_layers and n_ctx.

Node.js

npm install mullama. A native addon, not an HTTP client — the model lives in your Node process.

Go

go get github.com/cognisoc/mullama. In-process inference with ModelParams / ContextParams.

PHP

composer require mullama/mullama. Load a model and generate from PHP without a sidecar daemon.

C / C++

Link the static C ABI library directly — mullama_load_model, mullama_new_context, mullama_generate.

// compatibility

Drop-in where it counts

Ollama CLI

Same verbs: run, pull, serve, chat, list, ps, create, show, rm, cp. Muscle memory transfers.

Modelfile

The same Modelfile format and model registry — existing Modelfiles work unchanged.

Port 11434

Serves on the Ollama port, so existing client code pointed at localhost:11434 keeps working.

OpenAI API

OpenAI-compatible surface at /v1 — the openai SDK, LangChain, and LlamaIndex just work.

Anthropic API

Anthropic-compatible at /v1/messages — the anthropic SDK works too. Ollama has no equivalent.

Web UI / TUI

A built-in web UI and terminal UI ship with the runtime.

// deployment

Runs where you run

Embed in-process

Link the static C ABI into a desktop, CLI, edge, or mobile build — no daemon, no network.

7 GPU backends

CUDA, Metal, ROCm, Vulkan, OpenCL, SYCL, and RPC for distributed inference. CPU out of the box.

Apple Silicon

Metal is selected automatically on Apple Silicon; no configuration required.

Containers

ghcr.io/cognisoc/mullama for CPU and :<ver>-cuda for NVIDIA.

Production knobs

Per-model resource limits, LRU model eviction, persistent stats via sled, Prometheus metrics.

Static binaries

A shell one-liner pulls a static binary for Linux and macOS.

// capabilities

What the runtime can do

GGUF models

Any GGUF file from Hugging Face, with aliases for Llama, Qwen, DeepSeek, Mistral, Phi, Gemma.

Multimodal

Vision via LLaVA, llava-phi3, and Moondream; the model resolves from a short alias.

Audio

Real-time audio capture with voice activity detection and streaming responses.

Embeddings

Built-in embeddings via BGE and Nomic for RAG pipelines.

Grammar output

Grammar-constrained JSON output for structured generation.

Parallel batch

Work-stealing scheduling for parallel inference across documents (parallel feature).