language, license, tags, base_model, quantized_by, pipeline_tag

language	license	tags	base_model	quantized_by	pipeline_tag
en fr multilingual	apache-2.0	gguf quantized mac apple-silicon local-inference worthdoing	microsoft/Phi-3.5-mini-instruct	worthdoing	text-generation

Author: Simon-Pierre Boucher

Phi-3.5-mini-instruct - GGUF Quantized by worthdoing

Quantized for local Mac inference (Apple Silicon / Metal) by worthdoing

About

This is a GGUF quantized version of Phi-3.5-mini-instruct, optimized for running locally on Apple Silicon Macs with llama.cpp, Ollama, or LM Studio.

• Original model: microsoft/Phi-3.5-mini-instruct
• Parameters: 3.8B
• Quantized by: worthdoing
• Pipeline: corelm-model v1.0

Description

Microsoft's Phi 3.5 mini. Strong reasoning in a compact form factor.

Available Quantizations

File	Quant	BPW	Size	Use Case
phi-3.5-mini-instruct-Q4_K_M-worthdoing.gguf	Q4_K_M	4.58	~2.0 GB	Recommended - Best quality/size ratio
phi-3.5-mini-instruct-Q5_K_M-worthdoing.gguf	Q5_K_M	5.33	~2.4 GB	Higher quality, still fast
phi-3.5-mini-instruct-Q8_0-worthdoing.gguf	Q8_0	7.96	~3.5 GB	Near-original quality

How to Use

With Ollama

# Create a Modelfile
cat > Modelfile <<'MODELEOF'
FROM ./phi-3.5-mini-instruct-Q4_K_M-worthdoing.gguf
MODELEOF

ollama create phi-3.5-mini-instruct -f Modelfile
ollama run phi-3.5-mini-instruct

With llama.cpp

llama-cli -m phi-3.5-mini-instruct-Q4_K_M-worthdoing.gguf -p "Your prompt here" -ngl 99

With LM Studio

1. Download the GGUF file
2. Open LM Studio -> My Models -> Import
3. Select the GGUF file and start chatting

Quantization Method

Our quantization pipeline (corelm-model v1.0) follows a rigorous multi-step process to ensure maximum quality and compatibility:

Step 1 — Download & Validation

• Model weights are downloaded from HuggingFace Hub in SafeTensors format (.safetensors)
• Legacy formats (.bin, .pt) are excluded to ensure clean, verified weights
• Tokenizer, configuration, and all metadata are preserved

Step 2 — Conversion to GGUF F16 Baseline

• The original model is converted to GGUF format at FP16 precision using convert_hf_to_gguf.py from llama.cpp
• This lossless baseline preserves the full original model quality
• Architecture-specific tensors (attention, FFN, embeddings, MoE routing) are mapped to their GGUF equivalents

Step 3 — K-Quant Quantization

• The F16 baseline is quantized using llama-quantize with k-quant methods
• K-quants use a mixed-precision approach: more important layers (attention, output) retain higher precision, while less sensitive layers (FFN) are compressed more aggressively
• Each quantization level offers a different quality/size tradeoff:

Method	Bits per Weight	Strategy
Q4_K_M	~4.58 bpw	Mixed 4/5-bit. Attention & output layers use Q5_K, FFN layers use Q4_K. Best balance of quality and size.
Q5_K_M	~5.33 bpw	Mixed 5/6-bit. Attention & output layers use Q6_K, FFN layers use Q5_K. Higher quality with moderate size increase.
Q8_0	~7.96 bpw	Uniform 8-bit. All layers quantized to 8-bit. Near-lossless quality, largest file size.

Step 4 — Metadata Injection

• Custom metadata is embedded directly in each GGUF file:
  • general.quantized_by: worthdoing
  • general.quantization_version: corelm-1.0
• This ensures full traceability and provenance of every quantized file

Tools & Environment

• llama.cpp: Used for both conversion and quantization — the industry-standard open-source LLM inference engine
• Target platform: Apple Silicon Macs (M1/M2/M3/M4) with Metal GPU acceleration
• Inference runtimes: Compatible with llama.cpp, Ollama, LM Studio, koboldcpp, and any GGUF-compatible runtime

Recommended Hardware

Quant	Min RAM	Recommended
Q4_K_M	4 GB	Mac with 8 GB+ RAM
Q5_K_M	4 GB	Mac with 8 GB+ RAM
Q8_0	4 GB	Mac with 8 GB+ RAM

Tags

general, reasoning, coding, math

---

Quantized with corelm-model pipeline by worthdoing on 2026-04-17