Repository · read 2026-10-17
muse-spark-open-2x-dgx-spark
Documents the smallest NVIDIA pair that can hold Muse Spark Open: two DGX Spark nodes splitting a 194.6 GiB IQ1_M quant over llama.cpp RPC, with an OpenAI-compatible server, working tool calls and reasoning.
The repository → 0 stars · Python
| Field | Value |
|---|---|
| Repository | jlacroix82/muse-spark-open-2x-dgx-spark 0 stars · read 2026-10-17 |
| Language | Python Last push 2026-10-16 |
| Licence | Apache-2.0 Repository licence; the model weights carry their own. |
| Stack | llama.cpp RPC, CUDA 13, two GB10 nodes with about 121 GB unified memory each What it runs on. |
| Best for | Anyone pricing a local deployment on NVIDIA hardware who wants a measured starting point instead of a vendor claim. A reader, not a benchmark. |
How it works
The head node loads the GGUF and pushes tensors to the peer's rpc-server; each node holds roughly 97 GiB of weights. The repository pins the llama.cpp commit it measured against, now that the architecture support has merged upstream in build b11671.
What we checked
- The README's hardware requirement and the arithmetic behind two nodes
- The llama.cpp PR #30616 merge note and the b11671 prerelease
- The stated decode and prefill figures in the repository description
- The warning that rpc-server takes no model argument
What we did not check
We did not run the recipe; it needs two DGX Sparks and a 195 GB download that we do not have here.
Our note
A measured recipe for the smallest NVIDIA box that holds the model: two DGX Sparks, a 194.6 GiB IQ1_M quant, llama.cpp RPC over the 200 GbE fabric, and roughly seven tokens per second.
github.com/jlacroix82/muse-spark-open-2x-dgx-spark Verified 2026-10-17 Stars 0 at read time