By Hamza Faisal

A 284-billion-parameter AI model that once needed a rack of GPUs can now run on an ordinary desktop with 64GB of memory and no graphics card at all. The open-source project llama-modchip, released by developer Synchopate, compresses DeepSeek's V4-Flash model small enough to fit entirely in system RAM — and it has quickly become one of the hottest projects on Hacker News.

DeepSeek-V4-Flash is a mixture-of-experts model with 284.33 billion parameters in total, though only 13.17 billion are active for any given token. The standard low-bit build of the model weighs 76.87 GiB — too large for a 64GB machine before a single token of context is loaded. Llama-modchip transcodes it into a 56.56 GiB checkpoint at 1.71 bits per weight, using a custom "Fractal-64" tensor format, so the model plus a 256,000-token context window peaks at 57.61 to 58.9 GiB of resident memory.

Fast enough to use, barely enough room to breathe

On the developer's test machine — a 16-core AMD Ryzen 9 9955HX with 64GB of DDR5 and no GPU — prompt processing runs at 44.36 tokens per second, up 77 percent from the project's earlier baseline. Plain text generation runs at 9 to 10 tokens per second, rising to 14.7 tokens per second with speculative drafting switched on. That is slow next to a GPU, but fast enough for a local assistant, coding help, or long-document work on hardware many people already own.

The fit is tight. A machine sold as 64GB has roughly 59.6 GiB of usable capacity, so the model leaves as little as 0.7 GiB for the operating system and everything else running on it. The project also requires a CPU with AVX-512BW support, and its custom tensor type means stock llama.cpp, Ollama and LM Studio refuse to load the files — users need the project's own launcher or prebuilt server binary.

Why it matters beyond one model

The same release also packages smaller models for everyday machines: Nvidia's 30-billion-parameter Nemotron-3.5-30B-A3B runs at 25 to 111 tokens per second on systems with 16 to 32GB of memory. Together, the numbers sketch a future in which frontier-class models are no longer confined to data centres — they can live on a workstation, a laptop, or a home server, running entirely offline.

There are caveats. Every benchmark so far comes from the developer's own hardware, and no independent reproduction of the 284-billion-parameter result has been published yet. The project has nevertheless gathered thousands of GitHub stars within days, and efficient-inference researchers are already contributing. If the numbers hold up under outside testing, the barrier to running a genuinely large model at home just fell by an order of magnitude.