Modelpublic

webmp3/Sakura-Micro-Bonsai-2-GSQ-RCO

sourceHugging Faceapache-2.0updated 5d agoView on Hugging Face

0likes336downloads

Model Card

Sakura Micro Bonsai 2 GSQ-RCO

Sakura Micro Bonsai 2 GSQ-RCO is an experimental ultra-low-bit derivative of Prism ML's Ternary Bonsai 2 27B, itself derived from Qwen3.8-27B. This release applies GSQ/RCO-based mixed-codec allocation and additional low-bit compression.

Provenance / Base model

Sakura Micro Bonsai 2 GSQ-RCO is derived from Prism ML's Ternary Bonsai 2 27B, which identifies Qwen3.8-27B as its base model.

Release artifact

Sakura-Micro-Bonsai-2-GSQ-RCO-5.3GiB.gguf

—Size: 5,686,593,888 Bytes (5.296 GiB)
—SHA-256: 5054d9b3b3b95145ebf1def33803fd93a1962dda37ca50bd53dbcbbbe4655eb3
—The bundled SHA256SUMS file can be used to verify the downloaded GGUF.

Benchmark comparison

Metric	Ternary Bonsai 2 PTQ1 reference	Sakura Micro Bonsai 2 GSQ-RCO
File size	5.538 GiB	5.296 GiB
WikiText-2 PPL	10.2726	10.7251
Arithmetic	34/40 (85.0%)	37/40 (92.5%)
GSM8K	38/40 (95.0%)	35/40 (87.5%)
HumanEval	16/20 (80.0%)	15/20 (75.0%)
IFEval	9/12 (75.0%)	12/12 (100%)

Sakura Micro Bonsai 2 GSQ-RCO is 4.37% smaller than the 5,946,648,928-Byte Ternary Bonsai 2 PTQ1 reference. These are measurements on small benchmark subsets, not a claim of general quality superiority over the base model.

Compression approach

The project combines several ideas rather than applying one uniform quantizer to every tensor. It uses GSQ/RCO-based mixed-codec allocation and additional low-bit compression to target a smaller release artifact while retaining measured functionality.

This card intentionally does not disclose per-tensor selections, allocation details, internal optimization logic, or reproduction procedures.

Notes

—Experimental low-bit release.
—Use the runtime documented by the base model.

License

Apache-2.0. The license text is included in LICENSE.