A 4-bit oQ4h quantization of RadixArk/DaisyRose-27B for Apple Silicon, preserving the base model's multi-token-prediction (MTP) head for speculative decoding. The base model is released in BF16 (52 GB), FP8 (29 GB) and NVFP4 (23 GB); this repository is a much lighter MLX-native build at ~16 GB total (average 5.0 bits per weight).
| Property | Value |
|---|---|
| Architecture | Qwen3_5ForConditionalGeneration: 64 decoder layers in a 3:1 pattern of Gated DeltaNet linear-attention and full-attention layers, hidden size 5,120, 27-layer vision encoder, 1 multi-token-prediction layer |
| Parameters | 27.8B total (incl. 0.46B vision encoder and 0.42B MTP head) |
| Context length | 262,144 tokens |
| Input modalities | Text, images |
| Output modality | Text |
| Runtime | OMLX on Apple Silicon |
Weights use OMLX oQ4h mixed-precision affine quantization at an average of 5.0 bits per weight. Most linear layers are quantized at 4 bits / group size 64, with higher-bit allocations for sensitive modules:
| Module | Bits | Group size |
|---|---|---|
| Default (most linear layers) | 4 | 64 |
language_model.model.embed_tokens |
5 | 64 |
language_model.lm_head |
6 | 64 |
language_model.mtp.layers.0.self_attn.o_proj |
5 | 64 |
language_model.mtp.layers.0.self_attn.k_proj / v_proj |
8 | 64 |
The quantization was imatrix-calibrated on the oqe_code_multilingual code dataset (128 samples × 512 tokens); see oq_imatrix_report.json in this repository for the full calibration report.
No new evaluations have been run for this quantization. See the DaisyRose-27B model card for the base model's benchmark results (Terminal Bench, SWE Bench Pro v2, DeepSWE, HLE, IFBench, GPQA Diamond).
Apache-2.0, matching the base model's license.
model-00003-of-00004.safetensors 5.2 GBmodel-00002-of-00004.safetensors 5.1 GBmodel-00004-of-00004.safetensors 4.1 GBmodel-00001-of-00004.safetensors 3.2 GBRadixArk/DaisyRose-27B