CoolFace
Modelpublic

XxACCOxX/gemma3.5-48b_a4b

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
1likes61downloads
Model Card

Gemma3.5-48B-A4B-Q4KM

by XxACCOxX

Gemma3.5-48B-A4B-Q4KM is a community-built GGUF release that combines the instruction behavior of Gemma 3 27B with the Gemma 4 26B A4B MoE backbone through full-layer activation-distilled donor experts.

The model keeps the original Gemma 4 expert bank intact, appends a second donor bank derived from Gemma 3 across all 30 language layers, and preserves a Gemma 4-compatible inference path for local deployment.

Model Summary

  • —Formal name: Gemma3.5-48B-A4B-Q4_K_M
  • —Architecture: gemma4
  • —Quantization: Q4_K_M
  • —Total parameters: 48.1B
  • —Estimated active parameters: ~3.8B
  • —Context length: 262,144
  • —Total experts: 256
  • —Experts used per token: 8
  • —Expert layout:
  • —slots 0..127: original Gemma 4 experts
  • —slots 128..255: Gemma 3 activation-distilled donor experts

Construction

This model uses Gemma 4 26B A4B as the base MoE runtime and appends a full 30-layer donor expert bank derived from Gemma 3 27B.

The donor side was built through activation distillation rather than direct dense replacement. Non-expert backbone tensors remain aligned with the Gemma 4 runtime layout, while the appended donor experts extend the expert bank without overwriting the original Gemma 4 experts.

Previous release from the same author:

MMLU-Pro No-Think Result

On a fixed stratified MMLU-Pro subset with 280 items, 20 items per category across 14 categories, seed = 42, think = false, and num_predict = 2048, the model reaches:

  • —Gemma3.5-48B-A4B-Q4_K_M: 63.21% (177 / 280)
  • —Gemma 4 26B A4B base: 55.71% (156 / 280)

This is a +7.50 point improvement over the original Gemma 4 26B A4B base under the same no-think configuration.

Intended Use

Gemma3.5-48B-A4B-Q4KM is intended for local instruction-following use in direct-answer mode, with strong emphasis on general chat, mathematics, technical prompts, and broad knowledge tasks while retaining a relatively small active path compared with its total parameter count.

Format

This release is provided as a GGUF model for local inference stacks such as llama.cpp and Ollama-based workflows.

License

This release is derived from both Gemma 3 and Gemma 4 upstream checkpoints.

Gemma 4 model pages are published under Apache 2.0, while Gemma 3 model access remains tied to Google's usage license. Because this release combines both upstream lines, it should not be represented as a pure Apache-2.0 model artifact.

Use and redistribution should follow the applicable upstream terms for both source model families.