Measured Muse Glimmer recipe for DGX Spark

# Muse Glimmer on DGX Spark

A pinned, measured llama.cpp recipe for Meta's Muse Glimmer 30B on one NVIDIA GB10.

This repository downloads Meta's two official K-Quant targets, the official DFlash drafter, and the perception projector. It builds the exact llama.cpp revision used for the measurements, verifies every model file by SHA-256, serves an OpenAI-compatible API, and reproduces text, tool-call, coding, vision, and throughput checks.

> Muse Glimmer launched on August 10, 2026. Its llama.cpp support was still in PR [#26841](https://github.com/ggml-org/llama.cpp/pull/26841) when this recipe was measured, then merged as `62bf73d25c53b8161f8a22894d4f90c4aebbd7d0`. The scripts retain the exact exercised pre-merge commit so the published numbers remain reproducible.

## What is pinned

| Component | Identity |
|---|---|
| Hardware | NVIDIA GB10, 128 GB unified LPDDR5x |
| Model repo | `meta-models/Muse-Glimmer-30B-GGUF` |
| Model revision | `93769bc7ab5ad1e9cd22d857e3138cf5d977ae81` |
| llama.cpp | PR 26841, `f5902cbb4e620cc61d186dd6eb651e01d2363c8d` |
| CUDA target | `121`, resolved by llama.cpp to `121a` |
| Context used | 32,768 tokens |
| KV cache | F16 target and draft |
| Parallel slots | 1 |

Meta describes Muse Glimmer as a dense 29.6B-parameter agentic model with text and image input, 131K+ context, controllable reasoning, and a 1.8B-parameter perception encoder. The official GGUF release includes two main-model quantizations, a DFlash speculative drafter, and a multimodal projector.

## Quick start

Prerequisites on DGX OS:

```bash
sudo apt-get update
sudo apt-get install -y git cmake g++ curl libcurl4-openssl-dev
```

Docker is also required for the generated-code canary. The sandbox runs unprivileged with no network, a read-only root, dropped capabilities, and bounded CPU, memory, and process count.

Clone this recipe, then download and verify all four official artifacts:

```bash
git clone https://github.com/sxuff/muse-glimmer-dgx-spark
cd muse-glimmer-dgx-spark
./scripts/download-model.sh
```

Build llama.cpp for SM121:

```bash
./scripts/build-llama.sh
```

Start the higher-fidelity Dynamic quant with ordinary decoding:

```bash
./scripts/serve.sh dynamic baseline
```

The server binds to `127.0.0.1:8080` by default. Verify it:

```bash
curl -s http://127.0.0.1:8080/v1/models

curl -s http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content