---
license: gemma
library_name: gguf
base_model: google/gemma-4-26B-A4B-it
base_model_relation: quantized
model_name: Gemma-4-26B-A4B-it-Cerebellum-v6.1-templatefix-GGUF
model_creator: google
model_type: gemma4
quantized_by: deucebucket
pipeline_tag: text-generation
tags:
  - GGUF
  - gemma4
  - gemma
  - google
  - quantized
  - cerebellum
  - imatrix
  - moe
  - 3-bit
  - templatefix
model-index:
- name: Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF
  results:
  - task:
      name: Text Generation
      type: text-generation
    dataset:
      name: AI2 Reasoning Challenge (25-Shot)
      type: ai2_arc
      config: ARC-Challenge
      split: test
      args:
        num_few_shot: 25
    metrics:
    - name: normalized accuracy
      type: acc_norm
      value: 0.9556
    source:
      name: Local audited benchmark run (RTX 3090, llama.cpp)
      url: https://huggingface.co/deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF/tree/main/benchmark_results
  - task:
      name: Text Generation
      type: text-generation
    dataset:
      name: HellaSwag (10-Shot)
      type: hellaswag
      split: validation
      args:
        num_few_shot: 10
    metrics:
    - name: accuracy
      type: acc
      value: 0.8455
    source:
      name: Local audited benchmark run (RTX 3090, llama.cpp)
      url: https://huggingface.co/deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF/tree/main/benchmark_results
  - task:
      name: Text Generation
      type: text-generation
    dataset:
      name: MMLU-Redux (5-Shot)
      type: cais/mmlu
      config: all
      split: test
      args:
        num_few_shot: 5
    metrics:
    - name: accuracy
      type: acc
      value: 0.7133
    source:
      name: Local audited benchmark run (RTX 3090, llama.cpp)
      url: https://huggingface.co/deucebucket/Gemma-4-26B-A4B-it-Cerebellum-v6-GGUF/tree/main/benchmark_results
---

# Gemma 4 26B-A4B-it Cerebellum GGUF

This repository contains GGUF builds derived from
`google/gemma-4-26B-A4B-it`.

## 2026-05-22 Update

Added:

```text
gemma-4-26B-A4B-it-cerebellum-v6.1-templatefix.gguf
sha256: d24229facdef8360a7ffa8b37a50e1de636b9139a5eba0efe899828e45ae7989

gemma-4-26b-a4b-it.mmproj.gguf
sha256: b762c43119ebdc3e3c36d929d958e827fac35b03278dda9203f87131aee1f185
```

The v6.1 file keeps the v6 tensor allocation and updates GGUF/runtime-facing
metadata for Gemma 4 chat-template use. The update was tested with
`llama-server --jinja --reasoning auto` and request-level no-thinking controls.

Older files in this repository are retained for reproducibility.

## Tested Runtime

Runtime used for the 2026-05-22 templatefix checks:

```text
llama.cpp fork: https://github.com/deucebucket/llama.cpp
branch: cerebellum/gemma4-runtime-fixes
fork commit: ded491334 fix: harden Gemma 4 server budgets
base build: b8930-59fa0b455
```

Server shape used locally:

```bash
llama-server \
  --model gemma-4-26B-A4B-it-cerebellum-v6.1-templatefix.gguf \
  --mmproj gemma-4-26b-a4b-it.mmproj.gguf \
  --n-gpu-layers 99 \
  --ctx-size 65536 \
  --parallel 1 \
  --flash-attn on \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --jinja \
  --reasoning auto \
  --media-path /tmp/
```

Normal no-thinking requests used:

```json
{
  "chat_template_kwargs": {"enable_thinking": false},
  "thinking_budget_tokens": 0
}
```

Bounded-thinking smoke requests used `thinking_budget_tokens: 128`.

## 2026-05-22 Templatefix Test Artifacts

Creative-writing smoke files:

```text
creative_eval_20260522/regular_v6_1_templatefix_creative_summary.json
creative_eval_20260522/regular_v6_1_templatefix_creative_rerun_longcaps_summary.json
```

Non-coding tool-use files:

```text
agentic_eval_20260522/README.md
agentic_eval_20260522/regular_v6_1_noncoding_agentic_tools_strict_summary.json
```

Observed 2026-05-22 results from those artifacts:

| Area | Harness | Observed result |
|---|---|---|
| No-thinking output channel | six creative prompts | `reasoning_len=0` in recorded outputs |
| Template leakage markers | six creative prompts | no `<think>` marker or template marker recorded by checker |
| Creative long-cap rerun | four prompts rerun after initial length caps | four stop finishes in rerun summary |
| Non-coding tool workflow | three strict OpenAI-style tool tasks | `schedule_strict`, `release_notes_strict`, `creative_brief_strict` listed in `pass_cases` |

The non-coding tool harness used mock tools named `list_calendar`,
`create_calendar_hold`, `search_notes`, `save_note`, and `add_task`. It did not
test code editing.

## Evaluation

Benchmark results for the Cerebellum v6 tensor allocation, measured directly
on the GGUF with llama.cpp `llama-server` on an RTX 3090. The v6.1
templatefix file keeps the v6 tensor allocation with zero tensor changes
(metadata-only update), so these measurements describe the same weights.
Summary JSONs are in `benchmark_results/` in this repository.

| Benchmark | Cerebellum v6 (11 GB) | Local Q3_K_M baseline |
|-----------|:---:|:---:|
| ARC-Challenge | **95.56%** (1172 q) | 95.22% |
| HellaSwag | 84.55% (10042 q) | 86.57% |
| MMLU-Redux | 71.33% (2400 q) | 73.67% |

Protocol: multiple-choice benchmarks run against a local `llama-server` with
the project benchmark harness at temperature 0. HumanEval is not listed in
the metadata because the retained v6 HumanEval artifacts are marked for audit
in local notes. For Gemma 4, the current HumanEval/EvalPlus protocol uses the
chat-completions harness (`scripts/benchmark_evalplus_chat.py`) with
`enable_thinking: false`, `thinking_budget_tokens: 0`, and `BENCH_WORKERS=1`,
not raw completions.

## Historical Same-Repo Benchmark Artifacts

The following benchmark artifacts are from the earlier v6 line and the local
Q3_K_M baseline. They are included as historical same-project measurements, not
as new v6.1 measurements.

| Artifact set | ARC-Challenge | HellaSwag | MMLU-Redux | HumanEval note |
|---|---:|---:|---:|---|
| `q3km_baseline_*` | 95.2218 | 86.5664 | 73.6667 | `q3km_baseline_humaneval_results.json`: 62.2 pass@1 |
| `cerebellum_v6_*` | 95.5631 | 84.55 | 71.3333 | v6 HumanEval artifacts are retained but marked for audit in local notes |

For Gemma 4 HumanEval/EvalPlus, the local protocol now uses chat completions,
not raw completions:

```text
llama-server --jinja --reasoning auto
chat_template_kwargs: {"enable_thinking": false}
thinking_budget_tokens: 0
BENCH_WORKERS=1
```

## Files and Provenance

Main v6.1 GGUF:

```text
source base: google/gemma-4-26B-A4B-it
quantization family: mixed-precision GGUF
recipe lineage: Cerebellum v6 tensor allocation
base quant lineage: Q3_K_M with bartowski imatrix
```

Matching mmproj:

```text
gemma-4-26b-a4b-it.mmproj.gguf
```

## Notes

- The 2026-05-22 tests were run on local `llama-server`.
- The opencode coding-agent test is not used as a model-card result. In one
  internal White and Black project run, the model connected through the harness
  and ran a Godot test, then produced malformed edit-tool calls.
- The creative-writing checks are smoke tests plus mechanical checks, not a
  human preference benchmark.
- The non-coding tool checks use mocked tools and fixed task definitions.

## Credits

- Base model: Google Gemma Team, `google/gemma-4-26B-A4B-it`
- Imatrix source used in the v6 lineage: bartowski,
  `bartowski/google_gemma-4-26B-A4B-it-GGUF`
- GGUF/runtime: llama.cpp
- Method and quantization workflow: deucebucket/osmosis Cerebellum pipeline
- Local test artifacts: deucebucket Cerebellum workflow