Curiosity · AI Model
Phi-3.5 Mini
Phi-3.5 Mini is Microsoft Research's August 2024 tiny open-weights LLM — a 3.8B dense transformer trained on a high-quality filtered and synthetic curriculum. MIT-licensed and small enough to run on a modern smartphone via ONNX Runtime, it's the default starting point for on-device intelligent features.
Model specs
- Vendor
- Microsoft
- Family
- Phi
- Released
- 2024-08
- Context window
- 128,000 tokens
- Modalities
- text
- Input price
- $0.03/M tok
- Output price
- $0.03/M tok
- Pricing as of
- 2026-04-20
Strengths
- Open weights under MIT license — fully permissive commercial use
- 3.8B size — fits in ~2GB quantized, runs on phones
- 128K context window — unusual for tiny models
- Optimized for ONNX Runtime and DirectML on Windows
Limitations
- Small size limits complex reasoning and long-form generation
- Trails Gemma 2 9B and Llama 3.1 8B on many benchmarks
- Limited multilingual coverage relative to mid-tier models
- Best for narrow, well-specified tasks — not open-ended chat
Use cases
- On-device assistants for Windows, iOS, Android via ONNX Runtime
- Low-latency classification, routing, and intent detection
- Edge IoT inference on Raspberry Pi and similar devices
- Draft model for speculative decoding with larger targets
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU | ≈69% | 2024-08 |
| GSM8K | ≈86% | 2024-08 |
| HumanEval | ≈62% | 2024-08 |
Frequently asked questions
What is Phi-3.5 Mini?
Phi-3.5 Mini is Microsoft's 3.8B open-weights small language model released August 2024 under MIT license. Part of the Phi-3 family, it targets on-device AI with a heavy synthetic-data training curriculum.
Can Phi-3.5 Mini run on a phone?
Yes — quantized to INT4 it fits in ~2GB and runs on modern flagship phones via ONNX Runtime. That's the primary design target along with edge Windows devices.
Should I use Phi-3.5 Mini or Phi-4?
Phi-4 is much stronger on reasoning but 14B — it needs a real GPU. Phi-3.5 Mini is designed for phones and edge devices where 3.8B is the ceiling. Pick based on deployment target.
Sources
- Microsoft — Phi-3.5 family blog — accessed 2026-04-20
- Hugging Face — microsoft/Phi-3.5-mini-instruct — accessed 2026-04-20