Phi-3.5-mini punches way above its weight class — Microsoft's 3.8B model consistently beats models 2-3× larger on reasoning benchmarks. Q4_K_M keeps it under 2.5GB while preserving near-full quality. An ideal edge model when you need real intelligence with a small footprint.

Usage

llama.cpp (CLI)

bash

./llama-cli -m Phi-3.5-mini-instruct-LiteRT_Q4_K_M.gguf \
  -n 512 --temp 0.7 -p "<|system|>You are a helpful assistant.<|end|><|user|>"

Google AI Edge / MediaPipe (Android/iOS)

This GGUF is compatible with MLC-LLM and llama.cpp Android bindings for on-device inference. For use with Google Edge Gallery, convert to .task bundle using MediaPipe LLM conversion tools.

Python via llama-cpp-python

python

from llama_cpp import Llama

llm = Llama(
    model_path="Phi-3.5-mini-instruct-LiteRT_Q4_K_M.gguf",
    n_ctx=4096,
    n_threads=4,
    verbose=False,
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the derivative of sin(x²)?"},
    ]
)
print(response["choices"][0]["message"]["content"])

Ollama

ollama run hf.co/DuoNeural/Phi-3.5-mini-instruct-LiteRT

Performance Notes

Metric	Value
Quantization	Q4_K_M
RAM required	~3 GB (with context)
Recommended devices	6GB+ RAM phones, laptops
Quantization loss	Minimal — Phi-3.5 is robust to 4-bit quantization

Phi-3.5 Mini Highlights

3.8B params, trained on 3.4T tokens
Strong reasoning, coding, and instruction-following
128K context window (trimmed to device-safe lengths for edge)
One of the top 4B-class models in its generation

About the Conversion

Converted using llama.cpp GGUF pipeline with CUDA acceleration. Source weights downloaded from HuggingFace in safetensors format, converted to F16 GGUF, then quantized to Q4_K_M.

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

Platform	Link
HuggingFace	huggingface.co/DuoNeural
Website	duoneural.com
GitHub	github.com/DuoNeural
X / Twitter	@DuoNeural
Email	duoneural@proton.me
Newsletter	duoneural.beehiiv.com
Support	buymeacoffee.com/duoneural

DuoNeural Research Publications

Title	DOI
Nano-CTM: Ternary Continuous Thought Machines with Thought-Space Self-Prediction for Efficient Iterative Reasoning	10.5281/zenodo.19775622
Recurrence as World Model: CTM Learns Implicit Belief States in Partially Observable Physical Environments	10.5281/zenodo.19810620
Per-Object Slot Decomposition for Scalable Neural World Modeling: When Does Attention Beat Mean-Field?	10.5281/zenodo.19846804
The Dynamical Horizon Principle: CTM Gates Converge to the Predictability Limit of Dynamical Systems	10.5281/zenodo.19952612

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura — DuoNeural.

Research Team

Jesse — Vision, hardware, direction
Archon — Lab Director, post-training, abliteration, experiments
Aura — Research AI, literature synthesis, novel proposals

Subscribe to the lab newsletter at duoneural.beehiiv.com for model drops before they go anywhere else.