---
license: apache-2.0
base_model: meta-models/Muse-Glimmer-30B
library_name: transformers
pipeline_tag: text-generation
tags:
  - gptq
  - vllm
  - dflash
  - muse-glimmer
---

# Muse Glimmer DFlash assistant GPTQ Int4 G128

GPTQ 4-bit / group-128 / symmetric of Muse Glimmer's **5-layer DFlash drafter**. Apache 2.0, derived from Meta's DFlash assistant (not a standalone chat model).

Use as the speculative draft for [vLLM-XPU on one Arc Pro B70](https://github.com/mgaruccio/muse-glimmer-b70) with [`Muse-Glimmer-30B-GPTQ-Int4-sym-G128`](https://huggingface.co/mgaruccio/Muse-Glimmer-30B-GPTQ-Int4-sym-G128). Embeddings and lm_head are shared with the target at runtime.

- Quantizer: GPTQModel 7.3.2
- 35 decoder linears quantized; encoder / norms BF16
- vLLM fused module names in `quantize_config.json`: `qkv_proj`, `o_proj`, `gate_up_proj`, `down_proj`

## Experimental B70 concurrency sweep artifact

The C8–C128 short-burst sweep used this assistant with an experimental frozen
32,768-token vocabulary shortlist and DFlash K3. Download the public-safe
runtime input from [`artifacts/glimmer-b70-k3-shortlist-32768.json`](artifacts/glimmer-b70-k3-shortlist-32768.json).
It contains only the immutable original-vocabulary IDs and contract metadata—no
calibration prompts, votes, raw generations, or model weights.

This is **not** the default recipe or a production capacity claim. Its highest
observed short-burst median was 840.8 aggregate tok/s at C96; C48 was the
workload-specific latency/throughput knee. The measured workload used repeated
83-token prompts and capped 256-token reasoning-only outputs. See the
[presentation report](https://github.com/mgaruccio/muse-glimmer-b70/blob/38ea47b202135c22effa9d844ce73eb5603ab2a1/docs/concurrency-sweep.md)
for methodology, caveats, and the exact source patch at
[`b70-inference@faf4ba9`](https://github.com/mgaruccio/b70-inference/tree/faf4ba9889254719c878728fdd1a48806e7fd88b/scripts/experimental).

```bash
hf download mgaruccio/Muse-Glimmer-30B-assistant-GPTQ-Int4-sym-G128 --local-dir ./models/draft
```