janas
37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models c
What is janas?
37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.
Overview of janas
What it is
Janas is a family of tools for artificial intelligence written in C with no dependencies beyond the C library. Its first component, Janas-LLM, runs large mixture-of-experts language models on ordinary computers, on the CPU alone or with a GPU, and with the NPU of Intel's Core Ultra where there is one. Weights stay on disk, experts are streamed from an NVMe SSD, and frequently used experts stay in memory. An 80-billion parameter model can run on a laptop with 32 GB of RAM. It is described as an experiment, not a product, and is in early development.
Key features
- Runs large Mixture-of-Experts models (e.g., 80B parameters) on a 32 GB laptop by streaming expert weights from disk.
- Supports CPU-only, GPU (via Vulkan), and Intel NPU (Core Ultra) for reading prompts.
- Achieves higher read and write token rates than llama.cpp on all 13 models compared in Q4_K_M on the same laptop (Debian 13 Linux, Intel Core Ultra 9 185H, 32 GB).
- Auto-selects the fastest hardware path (CPU, GPU, NPU) per model and task within ~1.5% of the best benchmarked configuration.
- Written in pure C with no dependencies beyond the C library.
- Models are distributed as pre-converted
.jnsfiles (format version 2) from Hugging Face; format stability is guaranteed for future versions.
Use cases
- Running large language models (especially Mixture-of-Experts architectures) on ordinary consumer laptops that have less RAM than the model size.
- Comparing inference speed across different hardware backends (CPU, GPU, NPU) on the same machine.
- Experimenting with CPU-centric or hybrid CPU/GPU/NPU LLM inference without relying on cloud services.
Getting started
shcurl -LO https://github.com/prabanta-dev/janas/releases/latest/download/janas-linux-x86_64.tar.gztar xf janas-linux-x86_64.tar.gz && cd janas./janas-chatFor Windows, download janas-windows-x86_64.zip, unpack it, and in a terminal in the janas folder run .janas-chat.
License and pricing
Licensed under GPL-3.0. The project is open source.
Based on the project's public README and repository metadata.