Capability matrix
- Compute and format BF16 · Ryzen 9 9950X3D AVX-512 BF16 with oneDNN 3.12 and PyTorch SLEEF.
- Local entry point Specialized text runner.
- Server Recorded tool call. A 32K bundle completed a direct HTTP file-read and continuation.
- Harness Muse Code / Hermes pending. Captured Muse Code input exceeded 32K; tested Hermes required 64K. Retest with sufficient context.
Choose the download
Publisher model card · CKE input files
Publisher task: image text to text. CKE example: BF16. Use this artifact identifier:
hf://meta-models/Muse-Glimmer-30B.
Get started
BF16 AVX-512 + oneDNN lane; three 128-token trajectories are exact against PyTorch; vision and long context are not yet certified Use the model-specific setup.
What has actually run?
384/384 finite full-vocabulary logit rows and all greedy tokens bit-exact to pinned PyTorch eager BF16. Three 128-token trajectories; window-boundary and long-context certification remain open.
CPU evidence. Ryzen 9 9950X3D AVX-512 BF16 with oneDNN 3.12 and PyTorch SLEEF.
Performance. The cited model result does not give a comparable prefill/decode pair here. ARM throughput for this exact artifact is unverified here. Benchmark evidence · 32K to 128K results
Can I fine-tune it?
Full-checkpoint training, LoRA and other PEFT methods: no verified CKE recipe for this exact checkpoint in the reviewed evidence. For a trainable starter, use the FP32 and adapter recipes.