license: apache-2.0 base_model:
- Blackfrost-Research/Muse-Glimmer-30B-Abliterated-BF16 tags:
- muse-glimmer
- abliterated
- nvfp4
- quantized
- vllm pipeline_tag: image-text-to-text library_name: transformers
Refusal benchmark — R1-HARMFUL-BENCH-450
Measured on the abliterated model — quantization holds it, no refusal snapback:
| 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 |
Why this model exists
Muse Glimmer is Meta Superintelligence Labs' 30B agentic, on-device model. This is the abliterated build — refusal behavior removed via a Blackfrost weight-change process — packaged as NVFP4 (4-bit float) safetensors — ~28 GB for NVIDIA Blackwell GPUs. The local footprint is the product.
Specifications
| 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 |
Serving (vLLM — NVFP4, verified ~300 tok/s)
Proven fast path on NVIDIA Blackwell (RTX PRO 6000). NVFP4 (modelopt_fp4), tensor-parallel across 2 GPUs. --reasoning-parser muse_glimmer keeps the reasoning channel out of the answer text.
docker run --gpus all --network host --ipc host \
vllm/vllm-openai:muse-glimmer \
vllm serve Blackfrost-Research/Muse-Glimmer-30B-Abliterated-NVFP4 \
--served-model-name Muse-Glimmer-30B-Abliterated-NVFP4 \
--quantization modelopt_fp4 \
--tensor-parallel-size 2 \
--reasoning-parser muse_glimmer \
--trust-remote-code \
--max-model-len 8192 \
--gpu-memory-utilization 0.90 \
--host 0.0.0.0 --port 8000
OpenAI-compatible at http://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.
Sampling (Meta): 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.