APEX Quantization Series by IsValorum
[!IMPORTANT] Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.
If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and supports my university studies.
[!NOTE] Text GGUF. For multimodal use, grab
mmproj-ThinkingCap-gemma-4-26B-A4B-it-f16.gguffrom the official GGUF repo.MTP/DSpark: this conversion has no MTP tensors and no DSpark draft is included for this family.
The BF16 master GGUF of bottlecapai/ThinkingCap-gemma-4-26B-A4B-it, converted from the
official safetensors with llama.cpp's convert_hf_to_gguf.py (arch gemma4) and published as the
canonical source archive for the APEX-I V2.1 quantization recipes. No weights were quantized,
merged or modified: every tensor is either BF16 (266) or the source's F32 norms/scales/routers (392).
92b53c9b065d4710bbda9e5c84be751470386e7b (all 9 pinned files verified: size + sha256).cpu-performance, 2026-11-05.| Source file | Bytes | sha256 |
|---|---|---|
model-00001-of-00002.safetensors |
49,907,246,508 | 07e3e2f2…8a163 |
model-00002-of-00002.safetensors |
1,704,763,408 | 5ccdeb1a…9b4f2 |
tokenizer.json |
32,169,626 | cc8d3a0c…bfe0f |
| Output | Bytes | sha256 |
|---|---|---|
ThinkingCap-gemma-4-26B-A4B-it.BF16.gguf |
50,505,136,704 | 3a7755b9d9097719f8c6495ef8859863114c8fe3d71e991b8ee9d2fb8c403c28 |
token_embd.weight doubles as the output head.ffn_gate_inp, ffn_down_exps.scale, layer_output_scale) stay F32.The exact tensor breakdown below has been verified directly from the compiled binary weights:
| Layer Group | Sub-Component / Tensor | Qty | Precision | Engineering Rationale |
|---|---|---|---|---|
| Tied Output Head + Embeddings | token_embd.weight (tied) |
1 | BF16 |
Tied weights double as the output head; kept at full source precision. |
| All Normalizations & Scales | *_norm, *_scale, layer_output_scale, rope_freqs |
332 | F32 |
Uncompressed numerical stability, identical to the source checkpoint. |
| Expert Routers | ffn_gate_inp.weight, ffn_gate_inp.scale |
60 | F32 |
Routing over 128 experts stays at source precision. |
| Sliding Attention q/k/v | 25 sliding-window layers (window 1024) | 75 | BF16 |
Standard quadratic attention at full source precision. |
| Sliding Attention output | 25 sliding-window layers | 25 | BF16 |
Attention output projection over local context. |
| Full Attention Anchors q/k | layers 5, 11, 17, 23, 29 | 10 | BF16 |
Full-context anchor projections at source precision. |
| Full Attention Anchors output | layers 5, 11, 17, 23, 29 | 5 | BF16 |
Attention output projection over deep context. |
| Dense Shared FFN | ffn_{gate,up,down} (all 30 layers) |
90 | BF16 |
Per-layer foundation FFN active on every token. |
| Edge MoE Experts | Layers 0-9 & 20-29 (ffn_gate_up_exps, ffn_down_exps) |
40 | BF16 |
Routed experts at full source precision. |
| Core MoE Experts | Layers 10-19 (ffn_gate_up_exps, ffn_down_exps) |
20 | BF16 |
Deep-band routed experts at full source precision. |
Total: 658 tensors (266 BF16 + 392 F32). The machine-readable version is tensor-map.tsv.
| File | Size | Notes |
|---|---|---|
ThinkingCap-gemma-4-26B-A4B-it.BF16.gguf |
50,505,136,704 bytes | BF16 master GGUF (text) |
| tensor-map.tsv | per-tensor types | Name → precision map |
| chat_template.jinja | from source | ThinkingCap chat template |
llama-cli -m ThinkingCap-gemma-4-26B-A4B-it.BF16.gguf -p "Hello" -ngl 99 -c 32768
llama-server -m ThinkingCap-gemma-4-26B-A4B-it.BF16.gguf --jinja -ngl 99 -c 32768
Apache-2.0, as declared by the base model repository. See the source model card.