Views
No views yet
| Metric | Result |
|---|---|
| True refusal (harmful, n=300) | 0 / 300 = 0.0% |
| True refusal (full 450) | 0 / 450 = 0.0% |
| Substring-harmful | 0 / 300 |
| Substring-all | 2 / 450 (XSTest false positives) |
| Errors | 0 |
| Architecture | muse_glimmer — dense, 52 layers, hidden 6656, GQA (32 q / 2 kv), sliding-window attention, + vision tower |
| Base | meta-models/Muse-Glimmer-30B — Meta, Apache-2.0 |
| Transform | Abliterated — refusal behavior removed via a Blackfrost weight-change process; multimodal capability intact |
| Format | NVFP4 (4-bit float) safetensors — ~28 GB |
| Context | 131,072 |
modelopt_fp4), tensor-parallel across 2 GPUs. --reasoning-parser muse_glimmer keeps the reasoning channel out of the answer text.1docker run --gpus all --network host --ipc host \
2 vllm/vllm-openai:muse-glimmer \
3 vllm serve Blackfrost-Research/Muse-Glimmer-30B-Abliterated-NVFP4 \
4 --served-model-name Muse-Glimmer-30B-Abliterated-NVFP4 \
5 --quantization modelopt_fp4 \
6 --tensor-parallel-size 2 \
7 --reasoning-parser muse_glimmer \
8 --trust-remote-code \
9 --max-model-len 8192 \
10 --gpu-memory-utilization 0.90 \
11 --host 0.0.0.0 --port 8000http://localhost:8000/v1. Fits on a single 96 GB card (drop --tensor-parallel-size); TP=2 pushes aggregate throughput to ~300 tok/s (~80 tok/s single-stream). The chat template ships in the repo.temperature 1.0, top_p 0.95, top_k 64. It's a heavy thinker — use a generous max_tokens (≥ 1024) and steer depth with a Reasoning strength: low/medium/high/xhigh system line. Reasoning is returned separately from the final answer.