KaviarLabs/Shadow-Siren-26B-A4B-UD-Q4_K_XL-i1-GGUF

Shadow-Siren-26B-A4B — BF16 & UD-Q4_K GGUFs

GGUF conversions of Vortex5/Shadow-Siren-26B-A4B, published by KaviarLabs.

This repository contains a full BF16 GGUF plus three custom Unsloth-style variants: UD-Q4_K_S-i1, UD-Q4_K_M-i1, and UD-Q4_K_XL-i1. Each quant uses the exact tensor-by-tensor qtype allocation extracted from the corresponding Unsloth Gemma 4 26B-A4B GGUF together with the Shadow-Siren-specific i1 importance matrix published by mradermacher.

These are not official Unsloth quantizations. The UD-* names describe reproduced tensor precision maps; i1 identifies the Shadow-Siren imatrix source. The quantized weights are from Shadow-Siren.

Files

File	Description	Size
Shadow-Siren-26B-A4B-BF16.gguf	BF16 converted directly from source Safetensors	~47.04 GiB
Shadow-Siren-26B-A4B-UD-Q4_K_S-i1.gguf	UD-Q4_K_S map + Shadow-Siren i1 imatrix	15.355 GiB
Shadow-Siren-26B-A4B-UD-Q4_K_M-i1.gguf	UD-Q4_K_M map + Shadow-Siren i1 imatrix	15.784 GiB
Shadow-Siren-26B-A4B-UD-Q4_K_XL-i1.gguf	UD-Q4_K_XL map + Shadow-Siren i1 imatrix	15.843 GiB

No mmproj is included. Text generation works directly; vision use requires a compatible Gemma 4 projector separately.

Source model & upstream credits

Shadow-Siren was created by Vortex5 as a multi-stage merge for roleplay, creative writing, storytelling and conversational use. Its published recipe uses SAEF → Karcher → HCR and combines:

• zerofata/G4-MeroMero-26B-A4B
• ReadyArt/Serenity-26B-A4B
• google/gemma-4-26B-A4B-it
• Gryphe/Pantheon-Reasoning-26B-A4B-1.1
• ReadyArt/Dark-Scarlett-v1.0-26B-A4B
• Darkhn/Gemma-4-26B-A4B-Animus-V14.1-FFT

See the original Shadow-Siren model card for the authoritative merge recipe.

BF16 conversion

The source Safetensors were converted directly with mainline llama.cpp using convert_hf_to_gguf.py --outtype bf16. The output was validated as Gemma 4 GGUF V3 with 658 tensors, expected F32/BF16 source types, and the source chat template preserved.

Quantizer build used later:

llama.cpp build 10335 (74ce15741)
MSVC 19.51.36252.0 / Windows AMD64

Importance matrix

The quant uses Shadow-Siren-26B-A4B.imatrix.gguf from mradermacher/Shadow-Siren-26B-A4B-i1-GGUF. This imatrix was published specifically for Shadow-Siren; llama-quantize loaded 295 importance-matrix entries.

UD tensor maps

The qtype allocations were extracted tensor-by-tensor from the corresponding files in Unsloth's Gemma 4 26B-A4B GGUF repository. Shadow-Siren retains the same 658-tensor Gemma 4 26B-A4B topology, allowing each map to be transferred name-for-name and shape-for-shape.

Variant	F32	Q4_K	Q5_0	Q5_1	Q5_K	Q8_0	Dry-run	Final size
UD-Q4_K_S-i1	392	30	29	0	0	207	5.22 BPW	15.355 GiB
UD-Q4_K_M-i1	392	30	0	29	0	207	5.37 BPW	15.784 GiB
UD-Q4_K_XL-i1	392	29	0	29	1	207	5.39 BPW	15.843 GiB

Notable differences:

• S: blocks 0–28 use Q5_0 for ffn_down_exps.weight; all 30 ffn_gate_up_exps.weight tensors are Q4_K.
• M: blocks 0–28 use Q5_1 for ffn_down_exps.weight; all 30 ffn_gate_up_exps.weight tensors are Q4_K.
• XL: blocks 0–28 use Q5_1 for ffn_down_exps.weight; blk.29.ffn_gate_up_exps.weight is promoted to Q5_K while the other 29 remain Q4_K.
• All three keep blk.29.ffn_down_exps.weight in Q8_0 and token_embd.weight in Q8_0; norms/router/scales and other small tensors follow the F32 assignments from the respective source maps.

Quantization

Each variant was produced with llama-quantize, the Shadow-Siren i1 imatrix and a literal tensor-type file extracted from the corresponding Unsloth GGUF. Base qtypes were Q4_K_S for S and Q4_K_M for M/XL, using 16 threads.

Dry-run results:

BF16 model size: 48150.36 MiB (16.01 BPW)
UD-Q4_K_S:       15708.72 MiB (5.22 BPW)
UD-Q4_K_M:       16147.34 MiB (5.37 BPW)
UD-Q4_K_XL:      16207.84 MiB (5.39 BPW)

Final files:

UD-Q4_K_S-i1  16,487,610,816 bytes  15.355 GiB
UD-Q4_K_M-i1  16,947,542,464 bytes  15.784 GiB
UD-Q4_K_XL-i1 17,010,981,312 bytes  15.843 GiB

Validation

Every final GGUF was parsed and compared against its extracted Unsloth reference map tensor by tensor, requiring all 658 tensor names, shapes and qtypes to match exactly.

UD-Q4_K_S:  F32:392,Q4_K:30,Q5_0:29,Q8_0:207           exact_match=True
UD-Q4_K_M:  F32:392,Q4_K:30,Q5_1:29,Q8_0:207           exact_match=True
UD-Q4_K_XL: F32:392,Q4_K:29,Q5_1:29,Q5_K:1,Q8_0:207    exact_match=True

This proves exact reproduction of each qtype allocation map, not bit-identical weights with Unsloth's original files: the weights are Shadow-Siren and the imatrix is mradermacher's Shadow-Siren i1 matrix.

Usage

llama-server -m Shadow-Siren-26B-A4B-UD-Q4_K_XL-i1.gguf

Compatible GGUF frontends such as LM Studio can load the file as well.

Thanks

Full credit and thanks to Vortex5 (Shadow-Siren), Google (Gemma 4), zerofata (G4-MeroMero), ReadyArt (Serenity and Dark-Scarlett), Gryphe (Pantheon-Reasoning), Darkhn (Animus), mradermacher (Shadow-Siren i1 imatrix and GGUF work), Unsloth / Daniel Han (Dynamic/UD GGUF work and the referenced UD-Q4_K_XL allocation), ggml-org / llama.cpp contributors, and mergekit contributors.

Please support and consult the upstream repositories. This upload is not a replacement for their original model cards, licenses or documentation.

License

The source Vortex5/Shadow-Siren-26B-A4B repository is published under Apache-2.0. Users should also review applicable upstream terms before redistribution or deployment.