Muse Glimmer

Deployment

See the official SGLang installation guide for all installation methods and hardware platforms. The steps below match the Python and Docker options in the command panel.

pip install --upgrade pip
pip install uv
# Muse Glimmer support is not in a release yet -- build the PR branch:
# https://github.com/sgl-project/sglang/pull/34262
git clone -b muse-glimmer https://github.com/sgl-project/sglang.git
cd sglang
uv pip install -e "python[all]"

Run the Python output of the command panel below in that environment.

docker pull lmsysorg/sglang:dev-muse-glimmer

See Install → Method 3: Using Docker to start the image. Replace the inner sglang serve ... command with the command from the panel below.

Select a checkpoint format. Select whether to use speculative decoding:

• Standard: Use normal autoregressive decoding.
• DFlash: Use speculative decoding with the DFlash draft model. The draft serves as published, with no conversion step. See §2.

import { Deployment } from "/src/snippets/_deployment.jsx"; import { config } from "/src/snippets/configs/meta-models/muse-glimmer.jsx"; import { benchmarks } from "/src/snippets/configs/meta-models/muse-glimmer-benchmarks.jsx";

Playground

Use the Playground to test SGLang features that are not in the verified matrix. The Deploy panel above shows only combinations that the SGLang team has verified. The Playground lets you add more options to the command from the Deploy panel.

import { Playground } from "/src/snippets/_playground.jsx";

1. Model Introduction

Muse Glimmer is a multimodal reasoning model. You can serve Muse Glimmer in four formats:

• A BF16 checkpoint (MuseGlimmerForConditionalGeneration).
• A set of vendor GGUF files.
• A ready-to-serve NVFP4 + MXFP8 checkpoint.
• Three MLX repacks for Apple Silicon.

Resources: Muse-Glimmer-30B (BF16) · Muse-Glimmer-30B-assistant (DFlash draft) · Muse-Glimmer-30B-GGUF · Muse-Glimmer-NVFP4 · MLX · q4 · q4km-gs128 · q4k-dynamic.

2. Configuration Tips

The GGUF format is text only. SGLang has no mmproj path. You cannot use the vision GGUF files. Use the BF16 checkpoint for multimodal input.

The NVFP4 checkpoint. RadixArk/Muse-Glimmer-NVFP4 is a ready-to-serve NVFP4 + MXFP8 checkpoint. No conversion needed — point --model-path straight at it.

The DFlash draft. meta-models/Muse-Glimmer-30B-assistant is the vendor's native draft export and serves directly. No conversion needed.

DFlash with a GGUF target model needs --speculative-draft-load-format auto. Without this flag, the draft model uses the gguf load format from the target model. The loader then rejects the draft directory.

Apple Silicon uses an MLX checkpoint, not the GGUF files. The MLX backend has no GGUF path. Serve one of the three RadixArk/Muse-Glimmer-*-MLX artifacts with SGLANG_USE_MLX=1 (see the Apple Silicon cells in the command panel). All three take the same flags; q4km-gs128 is the one with a measured round. Keep --disable-radix-cache — the windowed KV storage for the sliding-window layers requires it — and set SGLANG_MLX_CACHE_LIMIT_GB=8 so the MLX buffer cache does not grow the footprint under concurrent load. Speculative decoding is not available on the MLX backend.

3. Advanced Usage

3.1 Reasoning

Muse Glimmer enables the muse reasoning parser by default. This parser separates the reasoning text from the final answer.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="meta-models/Muse-Glimmer-30B",
    messages=[{"role": "user", "content": "What is 15% of 240?"}],
)
msg = resp.choices[0].message
print("Reasoning:", getattr(msg, "reasoning_content", None))
print("Answer:", msg.content)

3.2 Tool Calling

Muse Glimmer enables the muse tool-call parser by default. This parser sends structured tool calls in message.tool_calls.

3.3 Multimodal

The BF16 checkpoint supports image input. It defaults to text only. To switch, select Modality in the command panel above.

Text only adds --language-model-only. This flag turns off the vision tower. SGLang does not build or load the vision weights. This frees memory for the KV cache. SGLang rejects image requests in this mode.

Select Image + text to turn on image input.

NVFP4, GGUF, and the MLX artifacts are text only. The Modality option does not appear for GGUF or MLX; NVFP4 only offers Text only.

3.4 Apple Silicon (MLX)

The MLX backend serves three Muse Glimmer artifacts on Apple Silicon Macs (48 GB unified memory or more). All three are text only — the MLX backend has no vision path — and all three take the same flags, so pick one in the command panel:

• RadixArk/Muse-Glimmer-q4-MLX — no measured round yet.
• RadixArk/Muse-Glimmer-q4km-gs128-MLX — a lossless repack of the vendor's Q4_K_M (gs128) GGUF: every weight keeps the GGUF's exact quantization code, with the group scales re-expressed in MLX affine bf16 (≤2⁻⁸ relative rounding). The numbers below are for this artifact.
• RadixArk/Muse-Glimmer-q4k-dynamic-MLX — no measured round yet.

Choose along the speed-versus-accuracy axis: footprint and expected accuracy both grow q4 → q4km-gs128 → q4k-dynamic, and decode speed moves the other way. Decode on Apple Silicon is memory-bandwidth-bound, so a smaller artifact reads fewer weight bytes per token — more tokens per second, and more unified memory left over for the KV cache. Take q4 for the fastest responses on the smallest machine, q4k-dynamic to stay closest to BF16, and q4km-gs128 for the middle ground — it is also the only one of the three with a measured round, below.

This table shows accuracy for the gs128 checkpoint, with the vendor llama.cpp fork serving the source GGUF on the same machine as the reference. GSM8K: 200 questions, no-thinking chat template, temperature 0, max 2048 new tokens. CIMemories: 1 profile, full combo, single trial, DeepSeek-R1-0528 judge.

CIMemories is a single-trial benchmark with a nondeterministic judge; treat the SGLang-vs-llama.cpp gap on that row as run noise, not a runtime effect. GSM8K parity is exact.

Decode throughput for gs128 on an M5 Pro (64 GB), 1k-in/1k-out greedy: 15.3 tok/s at batch 1, rising to 52.6 tok/s aggregate at batch 8 — ahead of llama.cpp on the same GGUF codes at every batch size above 1.