---

license: apache-2.0

language:

• en

base_model: Gryphe/Pantheon-Reasoning-26B-A4B-1.1-V2

base_model_relation: quantized

library_name: gguf

pipeline_tag: text-generation

tags:

• gguf
• llama.cpp
• quantized
• gemma4
• moe
• imatrix
• apex
• conversational
• roleplay
• reasoning
• creative-writing

---

Pantheon-Reasoning-26B-A4B-1.1-V2 — APEX GGUF

Mixed-precision APEX GGUF quantizations of

Gryphe/Pantheon-Reasoning-26B-A4B-1.1-V2.

These quants use a Pantheon-specific importance matrix and per-tensor mixed precision rather than a uniform GGUF quantization scheme.

SORRY, Im pretty sure this quantization is somewhat shit after some personal use? If not just comment BUT I basically went off and expected a quantization recipe of StyleTuneV2 to work well here but that model only had lm_head trained and this is vastly differently trained it seems so that was pretty stupid :P - sorry, my first quant

Available variants

| Variant | Description |

|---|---|

| APEX iQuality | Higher-fidelity mixed-precision quant. Approximately 5.89 BPW. |

| APEX iQuality — Q8 LM Head | Same iQuality tensor allocation, except output.weight is stored at Q8_0 for testing LM-head precision. |

| APEX iCompact | Smaller mixed-precision variant at approximately 4.31 BPW. |

Because APEX uses per-tensor mixed precision, the standard quant type at the end of each filename is an approximate size/BPW class rather than a description of every tensor in the file.

Fidelity benchmark

Measured against BF16 reference logits on WikiText-2 using llama.cpp, 128 × 512-token chunks (32,640 scored next-token positions):

| Variant | Mean KLD | Same top token | RMS Δp |

|---|---:|---:|---:|

| APEX iQuality | 0.861731 | 72.387% | 8.638% |

| APEX iQuality — Q8 LM Head | 0.856834 | 72.540% | 8.636% |

| APEX iCompact | 1.386163 | 61.759% | 11.638% |

The Q8 LM-head experiment produced only a small aggregate improvement over the standard iQuality variant.

Notes

Gemma 4 contains tensor dimensions that are incompatible with some requested K-quant formats. In particular, ffn_down_exps.weight has a 704-wide dimension, so llama.cpp may use compatible fallback formats when necessary.

Observed examples include:

• requested IQ4_XS → stored as IQ4_NL
• requested Q5_K → stored as Q5_1
• requested Q6_K → stored as Q8_0

These are expected shape-dependent fallbacks.

Base model

See the original model card for architecture, training details, intended usage, and model-specific notes:

Gryphe/Pantheon-Reasoning-26B-A4B-1.1-V2

Credits

Original model by Gryphe.

APEX mixed-precision methodology/configuration derived from the work by EmanuelOverride on Gemma 4 APEX GGUF quantization.