Brumby 14B Base — GGUF

GGUF quantizations of Manifest AI's Brumby 14B Base, a 14B parameter language model using power retention — a novel attention mechanism that replaces softmax attention with a degree-2 symmetric power feature map and linear recurrence.

What is Power Retention?

Power retention (paper) is a recurrent attention mechanism that:

• Expands keys and queries via a symmetric power feature map phi(x) (degree-2 polynomial expansion, head_dim → D=9216 for head_dim=128)
• Maintains a recurrent state S (updated as S = decay * S + outer(phi(k), v)) with a normalizer s
• Produces output as (phi(q) @ S) / (phi(q) @ s) — normalized linear attention

This gives the model constant-time per-token inference (like RWKV/Mamba) while maintaining strong language modeling performance.

Quantizations

How to Run

Important: This model requires a custom llama.cpp build with power retention support.

1. Build llama.cpp with power retention

git clone https://github.com/audreyt/llama.cpp
cd llama.cpp
git checkout power-retention
cmake -B build -DGGML_CUDA=ON  # or without CUDA for CPU-only
cmake --build build -j

2. Run inference

./build/bin/llama-cli \
  --model brumby-14b-base-pr.Q4_K_M.gguf \
  -p "Once upon a time" \
  -n 100 \
  --temp 0.7 \
  --repeat-penalty 1.2

Model Details

• Architecture: Qwen2PR (Qwen2 backbone with power retention replacing softmax attention)
• Parameters: 14.77B
• Context: 131,072 tokens
• Heads: 40 Q / 8 KV (GQA with 5 groups)
• Head dim: 128
• Layers: 40
• Vocab: 151,936 (Qwen2 tokenizer)
• Recurrent state: ~1.4 GB per sequence (9216 × 129 × 8 KV heads × 40 layers × 4 bytes)

Credits

• Model by Manifest AI
• Power retention algorithm from the retention package
• llama.cpp integration: audreyt/llama.cpp@power-retention