Muse Glimmer 30B R9V V1
Rough-draft release. This is the first public R9V model package, not the best available Muse Glimmer quant. It currently loses badly to Unsloth's higher-fidelity Q5/Q6 work in the quality evaluation used here. Do not infer quality superiority from the R9V name or the V1 label.
V1 is the public release name. V12 is the internal research lineage of the
exact GGUF and remains in manifests and hashes for reproducibility.
The complete package, including the optional projector and DFlash sidecar, is
pinned at Hugging Face revision
093f8ced7a8e2308b0f597084ebdbfa5f6614f75.
What it is
- 24,554,611,392-byte GGUF
- SHA256
f4870ff4ac316c1dbf50a55501f4c00e16070336fc40e119ff1167e43382856a - 731 tensors: 313 F32, 351 Q8_0, and 67 Q4_K
- seven gate/up pairs promoted to Q8_0 at layers 11, 15, 19, 23, 27, 31, and 35
- derived from Unsloth's pinned Q8_0 Muse GGUF with llama.cpp at the pinned
revision in
sources.lock.json
Quality disclosure
All figures below use the same native-BF16 teacher, evaluator, 480 chunks, 512 tokens per chunk, and 122,400 evaluated positions.
| Quant | Bytes | Mean KLD (lower is better) | Same top prediction |
|---|---|---|---|
| R9V V1 / V12 | 24,554,611,392 | 0.006121 | 96.879% |
| Unsloth Q4 comparison | 15,878,222,368 | 0.016883 | 94.806% |
| Unsloth UD-Q5_K_XL | 21,789,618,976 | 0.003071 | 97.724% |
| Unsloth UD-Q6_K_XL | 26,265,362,976 | 0.001034 | 98.752% |
On this evaluator V1 beats the compared Unsloth Q4, but that is not the useful headline: V1 has about 1.99x the KLD of Unsloth Q5 despite being roughly 2.76 GB larger, and about 5.92x the KLD of Unsloth Q6. This package is a speed-oriented engineering draft and a reproducible starting point for R9V, not a recommendation over Unsloth's quality quants.
Current product limitations
- The published R9V user runtime is not complete yet; current speed figures come from the frozen raw-token proof engine.
- No OpenAI-compatible API or chat template is qualified for this profile.
- Vision and DFlash artifacts are optional package components, but the speed comparison does not exercise them.
- The proof engine is specialized for gfx1201 and the exact tensor layout.
- The TG run reports 208 attention-pin fallbacks per 256-token sample; this is disclosed in the benchmark record rather than hidden.
See the R9V benchmark report and qualification report.
License and provenance
The model artifact remains under Apache License 2.0. It is derived
immediately from Unsloth's pinned Q8_0 GGUF of Meta's Muse Glimmer 30B model
and was requantized with llama.cpp/ggml tooling. Exact revisions and hashes are
in sources.lock.json; complete attribution and Meta
usage-policy guidance are in
THIRD_PARTY_NOTICES.md.