unsloth
Phi 4 mini reasoning
Phi 4 mini reasoning is a 3.84B dense phi3 model from unsloth. Configured context length is 131,072 tokens. License: mit. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters3.8B
Context128K
Attentiongqa
Licensemit
32 layers · 8 KV heads · head dim 128 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 7.15 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 4.74 GB | Hugging Face file bytes |
| Q8_0 | 3.80 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 3.30 GB | Hugging Face file bytes |
| Q6_K | 2.94 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 2.59 GB | Hugging Face file bytes |
| Q5_K_M | 2.65 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 2.29 GB | Hugging Face file bytes |
| Q4_K_M | 2.32 GB | Hugging Face file bytes |
| IQ4_NL | 2.17 GB | Hugging Face file bytes |
| IQ4_XS | 2.07 GB | Hugging Face file bytes |
| Q4_1 | 2.35 GB | Hugging Face file bytes |
| Q4_0 | 2.17 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 1.95 GB | Hugging Face file bytes |
| Q3_K_M | 1.97 GB | Hugging Face file bytes |
| Q3_K_S | 1.77 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 1.55 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 1.60 GB | Hugging Face file bytes |
| Q2_K | 3.13 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 1.20 GB | Hugging Face file bytes |
| UD-IQ1_M | 1.11 GB | Hugging Face file bytes |
| UD-IQ1_S | 1.06 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Tight fit | UD-Q8_K_XL | 26–35 |
| GeForce RTX 3060 12GB | Runs | BF16 | 23–31 |
| GeForce RTX 4060 Ti 16GB | Runs great | BF16 | 18–25 |
| GeForce RTX 4090 | Runs great | BF16 | 63–86 |
| GeForce RTX 5090 | Runs great | BF16 | 113–154 |
| Apple M4 Pro | Runs great | BF16 | 17–23 |
| Apple M5 | Runs | BF16 | 9.6–13 |
| CPU only · DDR5-5600 | Runs on CPU | BF16 | 5.6–7.7 |
Smallest devices that fit it fully
- GeForce RTX 3050 6GB — Tight fit, Q8_0, 20–27 tok/s
- GeForce RTX 2060 6GB — Tight fit, Q8_0, 40–55 tok/s
- GeForce GTX 1660 Super — Tight fit, Q8_0, 40–55 tok/s
- GeForce GTX 1660 Ti — Tight fit, Q8_0, 34–47 tok/s
- Intel Arc A380 — Tight fit, Q8_0, 22–30 tok/s
- GeForce RTX 5060 Ti 8GB — Tight fit, UD-Q8_K_XL, 40–55 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (BF16).
llama.cpp
llama-server -hf unsloth/Phi-4-mini-reasoning-GGUF:BF16 -c 8192 -ngl 32
Ollama
ollama run phi
LM Studio
lms get unsloth/Phi-4-mini-reasoning-GGUF # Then load quant BF16