Your models, on your hardware. Built from the kernel up.
Get early accessOn-device AI: models run on the machine in front of you — encrypted at rest, working with no connection at all. Your prompts, your files, and everything the system learns about how you work never need to leave it. Cloud models are a choice you make, never a dependency you inherit.
Nothing in Cosmos holds standing permission to send your data anywhere. Every data access is a capability — a cryptographic token that cannot be forged. Only you decide where your data goes.
A heterogeneous scheduler places work across CPU, GPU, and NPU wherever is most optimal. You already have the hardware to run AI, you just have to use it.
Cosmos currently supports Intel and AMD x86 PCs. Snapdragon (ARM) and RISC-V device support is coming soon. LLMs run on-device, 2.8× to 7.3× faster than llama.cpp on Linux on the same laptops. Nothing is sent to the cloud. We tested the latest Qwen models on consumer-grade laptops — see the comparisons below.
2.8×faster than on Linux
MSI Prestige 16 AI+ C3MTG · Core Ultra X7 358H · 32 GB
35B model, all the weights fit in memory.
4.4×faster than on Linux
MSI Prestige 16 AI+ C3MTG · Core Ultra X7 358H · 32 GB
27B dense model, all the weights fit in memory.
7.3×faster than on Linux
Lenovo Yoga 7 2-in-1 16ILL10 · Core Ultra 5 226V · 16 GB
35B model larger than the laptop's memory. Weights are streamed from the SSD.
![Left, the Cosmos chat running QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 at Xhigh effort: a Python test harness for a matmul function, then its actual output, the 2×2 product [[19, 22], [43, 50]], a 3×3 identity check and an associativity check both True, and timings for sizes 1 to 100 rising from 0.0 to 133.3 ms, followed by observations that the checks passed and the scaling is cubic. The status line reads 1039 tokens at 23.4 tok/s, context 5281 of 187392, 5.41 tokens a step. Right, a terminal running kore status mem: 26.7 of 30.8 GiB of physical memory in use, page classes of 4 KiB, 2 MiB and 1 GiB, and DRAM stream bandwidth per core class.](/_next/static/media/agent-gemm.1kshv5zq9k_n_.webp)
(open full size)Qwen3.8-27B writes matrix multiplication in Python, runs it with its own Python tool and checks it: a known product, the identity, associativity and cubic scaling. 23.4 tok/s with a 187K-token context, and no cloud anywhere in the loop.
Vision
So AI runs in someone else's data center, and your data goes along with every request. Your OS at home isn't designed to run fast AI.
Every popular OS predates AI by a generation. Everything was designed for a person clicking through files and apps, not for AI agents that read, decide, and act on their own.
Most machines sold in the last couple of years already have a CPU, a GPU, and an NPU — a dedicated AI chip built specifically for matrix math. They could run capable models locally, but most software still only ever asks one of them to do anything. When that's not enough, they ship the work to the cloud instead while your machine sits idle.
Cosmos treats all three as one machine to schedule across, not three separate afterthoughts — which is what makes a capable model practical on a laptop in the first place.
We intend to be the standard operating system for local AI.
Fast, secure, and private AI should belong to everybody. You are paying for your hardware, so you should get to use it.
Get early access
