Muse-Glimmer-30B-JANG_4M

JANG mixed-affine quantization of meta-models/Muse-Glimmer-30B for Apple Silicon runtimes that implement Muse Glimmer's dense multimodal architecture.

Bundle

Field	Value
Source revision	f84ecc3a0ea984a4c04542a84269e3d065350a6e
Format	jang_affine / JANG v2
Profile	JANG_4M
Effective language-weight bits	4.63
Indexed size	20.20 GB
Safetensor shards	32
Indexed tensor keys	2272
Language layers	52: 39 sliding + 13 full attention
Vision tensors	809 FP16 passthrough tensors
Assistant / DFlash	Not included; maintained as a separate artifact

This is a post-training quantization baseline. The upstream BF16 checkpoint contains no QAT weights or QAT scale metadata. GPTQ, imatrix, and AWQ were not applied, and the metadata says so explicitly.

Native model contract

• Text input and text output are part of the source architecture.
• The vision tower, adapter, projection, image processor, and video processor sidecars are preserved. Image/video execution in vMLX remains unverified.
• Reasoning uses reasoning_strength=low|medium|high|xhigh; omission defaults to high in the shipped chat template.
• Reasoning output is an assistant to=self channel. Visible content is addressed to=user.
• Tools use the ATEM function-call grammar. A Muse-specific incremental reasoning parser and ATEM tool parser are required.
• Deployment sampling follows the upstream model-card recommendation: do_sample=true, temperature=1.0, top_p=0.95, and top_k=64. BOS is 200000, EOS is [200001, 200008], pad is 200018, and maximum length is 131072.
• Cache topology is heterogeneous: rotating KV with window 2048 on 39 layers and unbounded KV on full-attention layers 3, 7, ..., 51.

Runtime status

PARTIAL / runtime unverified. Current validation covers source identity, safetensor headers, index integrity, mixed-bit metadata, exact processor/chat sidecars, deployment generation metadata, FP16 vision preservation, and selected dequantized-vs-BF16 tensor comparisons. It does not yet cover coherent generation in vmlx-swift, image grounding, video grounding, reasoning streaming, an ATEM tool round trip, multi-turn behavior, or prefix/partial-block/suffix cache reuse.

Do not interpret repository availability or structural loading as production readiness. This format is not a uniform mlx_lm quant; loaders must honor every per-module entry in config.json.quantization.

Files

• config.json: Muse architecture plus per-module affine overrides.
• jang_config.json: source revision, profile, capability, reasoning/tool, modality, generation, and mixed full/sliding cache metadata.
• chat_template.jinja: exact upstream Muse channel/ATEM template.
• processor_config.json: exact upstream image/video processor contract.
• generation_config.json: upstream-recommended deployment sampling plus the source token IDs and maximum length.
• LICENSE and USAGE_POLICY.md: copied from the pinned upstream source.

Download

hf download OsaurusAI/Muse-Glimmer-30B-JANG_4M \
  --local-dir ~/models/OsaurusAI/Muse-Glimmer-30B-JANG_4M

Verification

From the JANG repository:

PYTHONPATH=jang-tools uv run --no-project \
  --with mlx --with numpy --with safetensors --with tqdm \
  python jang-tools/scripts/verify_muse_glimmer_artifact.py \
  ~/models/OsaurusAI/Muse-Glimmer-30B-JANG_4M \
  --profile JANG_4M --dequant

Korean summary

이 번들은 공식 meta-models/Muse-Glimmer-30B BF16 체크포인트를 Apple Silicon용 JANG 혼합 affine 형식으로 변환한 JANG_4M PTQ 모델입니다. 52개 언어 레이어는 슬라이딩/전체 어텐션 구조를 유지하며, 비전 타워·어댑터·프로젝션 809개 텐서는 FP16으로 보존됩니다. 기본 추론 강도는 high이고 도구 호출은 ATEM 형식입니다. 현재 파일 구조, 메타데이터, 사이드카 및 일부 역양자화 비교는 확인했지만, vmlx-swift 실제 생성, 이미지/비디오, 추론 스트리밍, 도구 왕복 및 캐시 재사용은 아직 검증되지 않았습니다. 별도 5레이어 DFlash assistant는 포함하지 않습니다.

License and use

Apache 2.0 license and the upstream Muse Glimmer Usage Policy apply. Review LICENSE and USAGE_POLICY.md before use.

Contact

eric@osaurus.ai