ByteOtter / Gemma-4-E4B-IT-TAK-Reasoning-GGUF

Gemma 4 E4B IT — TAK reasoning winner

This repository contains the TAK reasoning winner for google/gemma-4-E4B-it . It is a reasoning-specialized, mixed-precision GGUF built under a fixed low-bit size budget. No fine-tuning or model merge was applied.

Part of the TAK Reasoning Champions collection.

File

Benchmark results

The headline category is reasoning. Other suites are diagnostics, not product claims. The matched low-bit comparator is Unsloth UD-IQ2_M , at 3.545 GB. TAK and Unsloth completed all 960 evaluation items with zero failed items. The BF16 source completed all 128 reasoning items with zero failed items.

Reasoning headline

The last column is TAK minus the listed artifact. Positive values favor TAK.

Full suite diagnostics

Scores are percentages from the same frozen MLab v1 evaluation pack. Full-suite deltas are TAK minus Unsloth in percentage points. The comparator artifact has SHA-256 ae7c1bd7867d2037c3d5b281c151319380a00f9f9f8de2cc8b507aef5d6a31b9 .

Process overview

1. Start from the upstream BF16 instruction checkpoint.
2. Build a reasoning-oriented calibration sample.
3. Evaluate mixed-precision candidates under a fixed byte ceiling.
4. Freeze the strongest validated candidate and run the full evaluation pack.

TAK changes weight precision. It does not fine-tune the model, merge weights, or replace the tokenizer or chat template. Internal scoring weights, search heuristics, candidate rules, and the tensor-level allocation recipe are not published.

Usage

Use a recent llama.cpp build:

llama-cli \
  -hf ByteOtter/Gemma-4-E4B-IT-TAK-Reasoning-GGUF \
  -p "Explain your answer carefully: If all A are B and no B are C, can any A be C?" \
  -n 512

Limitations

• This model targets reasoning. The off-target scores only describe diagnostic behavior.
• Quantization can change outputs, accuracy, and stability across prompts and runtimes.
• This repository contains the text-generation GGUF only. It does not include image or audio components.
• Review the base model card and validate the model for your application, especially for high-stakes use.

Related research

• Gemma 4 Technical Report
• You Had One Job: Per-Task Quantization Using LLMs’ Hidden Representations (TAQ)
• Beyond Activation Alignment: The Alignment-Diversity Tradeoff in Task-Aware LLM Quantization (TASA)

TAK was developed independently and is not an official implementation of TAQ or TASA.

Support

If this model is useful to you, you can support ByteOtter on Buy Me a Coffee . You can also follow ByteOtter on X .

License and attribution

This quantized derivative follows the base model’s Apache License 2.0 . See LICENSE and NOTICE for the included license and attribution notice.