phi-4-BonfyreFPQ3

BonfyreFPQ — BF16 safetensors format. Drop-in replacement, no special loader required.

Usage

Standard BF16 safetensors — load directly in PyTorch, diffusers, or any HuggingFace-compatible framework:

from safetensors.torch import load_file
weights = load_file("model.safetensors")

Or load with the transformers library as usual.

Compression Method

BonfyreFPQ v9/v10 with Bonfyre Weight Algebra:

1. Decompose W = L + R (truncated SVD)
2. Prune R with hybrid structure-aware pruning
3. Curl + divergence energy correction
4. FPQ v9 multi-scale encode (LR + E8 + RVQ + QJL + Ghost)
5. Decode back to BF16 safetensors

Quality

Per-weight cosine similarity ~0.9999 at ~4 bits/weight. See verified benchmarks.

Compressing Your Own Models

git clone https://github.com/Nickgonzales76017/bonfyre-oss.git && cd bonfyre-oss/cmd/BonfyreFPQ && make
./bonfyre-fpq algebra-compress input.safetensors output.safetensors --bits 3

Links

• BonfyreFPQ Source Code
• Build & Usage Guide
• All models
• Project Page