Microsoft · 15B · Q4_K_M · 8.43 GiB

Phi-4 at Q4_K_M needs about 11.0 GiB at 8K context, and the smallest listed machine that fits it is GeForce RTX 4060 Ti 16GB.

Dense 14B-class Phi-4. head_dim is hidden_size / num_attention_heads from config.json (5120 / 40).

ollama run phi4:14b

Weights: phi-4-Q4_K_M.gguf. Architecture: config source. Hugging Face repo created 2024-12-11. License field: mit.

Quants with a measured file size

QuantFileSize
Q4_K_M · recommendedphi-4-Q4_K_M.gguf8.43 GiB
Q5_K_Mphi-4-Q5_K_M.gguf9.88 GiB
Q6_Kphi-4-Q6_K.gguf11.2 GiB
Q8_0phi-4-Q8_0.gguf14.5 GiB

Machines

Same quant and 8,192 tokens on every row.
MachineMemoryTotalVerdictSpeed
Apple M1, 8 GB8 GB unified11.0 GiBNo—
Apple M2, 8 GB8 GB unified11.0 GiBNo—
Apple M3, 8 GB8 GB unified11.0 GiBNo—
GeForce RTX 3060 Ti8 GB GDDR11.0 GiBOffload4–6 tok/s
GeForce RTX 30708 GB GDDR11.0 GiBOffload4–6 tok/s
GeForce RTX 40608 GB GDDR11.0 GiBOffload4–6 tok/s
GeForce RTX 4060 Ti 8GB8 GB GDDR11.0 GiBOffload4–6 tok/s
GeForce RTX 50608 GB GDDR11.0 GiBOffload7–11 tok/s
GeForce RTX 5060 Ti 8GB8 GB GDDR11.0 GiBOffload7–11 tok/s
Radeon RX 66008 GB GDDR611.0 GiBOffload4–5 tok/s
Radeon RX 76008 GB GDDR611.0 GiBOffload4–6 tok/s
GeForce RTX 3080 10GB10 GB GDDR10.00 GiBTight43–67 tok/s
GeForce RTX 3060 12GB12 GB GDDR11.0 GiBTight18–32 tok/s
GeForce RTX 407012 GB GDDR11.0 GiBTight26–45 tok/s
GeForce RTX 4070 Super12 GB GDDR11.0 GiBTight26–45 tok/s
GeForce RTX 4070 Ti12 GB GDDR11.0 GiBTight26–45 tok/s
GeForce RTX 507012 GB GDDR11.0 GiBTight34–59 tok/s
Intel Arc B58012 GB GDDR611.0 GiBTight23–40 tok/s
Radeon RX 6700 XT12 GB GDDR611.0 GiBTight20–34 tok/s
Radeon RX 7700 XT12 GB GDDR611.0 GiBTight22–38 tok/s
Apple M1 Pro, 16 GB16 GB unified11.0 GiBTight10–18 tok/s
Apple M1, 16 GB16 GB unified11.0 GiBTight3–6 tok/s
Apple M2 Pro, 16 GB16 GB unified11.0 GiBTight10–18 tok/s
Apple M2, 16 GB16 GB unified11.0 GiBTight5–9 tok/s
Apple M3, 16 GB16 GB unified11.0 GiBTight5–9 tok/s
Apple M4, 16 GB16 GB unified11.0 GiBTight6–11 tok/s
GeForce RTX 4060 Ti 16GB16 GB GDDR11.0 GiBFits15–25 tok/s
GeForce RTX 4070 Ti Super16 GB GDDR11.0 GiBFits34–59 tok/s
GeForce RTX 408016 GB GDDR11.0 GiBFits37–63 tok/s
GeForce RTX 4080 Super16 GB GDDR11.0 GiBFits38–65 tok/s
GeForce RTX 5060 Ti 16GB16 GB GDDR11.0 GiBFits23–40 tok/s
GeForce RTX 5070 Ti16 GB GDDR11.0 GiBFits46–79 tok/s
GeForce RTX 508016 GB GDDR11.0 GiBFits49–85 tok/s
Intel Arc A770 16GB16 GB GDDR611.0 GiBFits26–45 tok/s
Legion Go16 GB LPDDR5X-750011.0 GiBTight6–11 tok/s
Radeon RX 6800 XT16 GB GDDR611.0 GiBFits26–45 tok/s
Radeon RX 7800 XT16 GB GDDR611.0 GiBFits32–55 tok/s
Radeon RX 9060 XT 16GB16 GB GDDR611.0 GiBFits16–28 tok/s
Radeon RX 9070 XT16 GB GDDR611.0 GiBFits33–57 tok/s
ROG Ally16 GB LPDDR5-640011.0 GiBTight5–9 tok/s
Steam Deck LCD16 GB LPDDR5-550011.0 GiBTight5–8 tok/s
Steam Deck OLED16 GB LPDDR5-640011.0 GiBTight5–9 tok/s
Apple M3 Pro, 18 GB18 GB unified11.0 GiBFits8–13 tok/s
Radeon RX 7900 XT20 GB GDDR611.0 GiBFits41–71 tok/s
Apple M2, 24 GB24 GB unified11.0 GiBFits5–9 tok/s
Apple M3, 24 GB24 GB unified11.0 GiBFits5–9 tok/s
Apple M4 Pro, 24 GB24 GB unified11.0 GiBFits14–24 tok/s
Apple M4, 24 GB24 GB unified11.0 GiBFits6–11 tok/s
GeForce RTX 309024 GB GDDR11.0 GiBFits48–83 tok/s
GeForce RTX 409024 GB GDDR11.0 GiBFits52–89 tok/s
Radeon RX 7900 XTX24 GB GDDR611.0 GiBFits49–85 tok/s
ROG Ally X24 GB LPDDR5X-750011.0 GiBFits6–11 tok/s
Apple M1 Max, 32 GB32 GB unified11.0 GiBFits21–35 tok/s
Apple M1 Pro, 32 GB32 GB unified11.0 GiBFits10–18 tok/s
Apple M2 Max, 32 GB32 GB unified11.0 GiBFits21–35 tok/s
Apple M2 Pro, 32 GB32 GB unified11.0 GiBFits10–18 tok/s
Apple M4, 32 GB32 GB unified11.0 GiBFits6–11 tok/s
GeForce RTX 509032 GB GDDR11.0 GiBFits92–158 tok/s
Apple M3 Pro, 36 GB36 GB unified11.0 GiBFits8–13 tok/s
Apple M4 Max 14-core, 36 GB36 GB unified11.0 GiBFits21–36 tok/s
Apple M3 Max 16-core, 48 GB48 GB unified11.0 GiBFits21–35 tok/s
Apple M4 Max 16-core, 48 GB48 GB unified11.0 GiBFits28–48 tok/s
Apple M4 Pro, 48 GB48 GB unified11.0 GiBFits14–24 tok/s
Apple M1 Max, 64 GB64 GB unified11.0 GiBFits21–35 tok/s
Apple M2 Max, 64 GB64 GB unified11.0 GiBFits21–35 tok/s
Apple M3 Max 16-core, 64 GB64 GB unified11.0 GiBFits21–35 tok/s
Apple M4 Max 16-core, 64 GB64 GB unified11.0 GiBFits28–48 tok/s
Apple M4 Pro, 64 GB64 GB unified11.0 GiBFits14–24 tok/s
Apple M2 Max, 96 GB96 GB unified11.0 GiBFits21–35 tok/s
Apple M3 Max 16-core, 128 GB128 GB unified11.0 GiBFits21–35 tok/s
Apple M4 Max 16-core, 128 GB128 GB unified11.0 GiBFits28–48 tok/s

Exact answers

GeForce RTX 3060 12GBtightGeForce RTX 4060offloadGeForce RTX 4060 Ti 16GBfitsGeForce RTX 4070tightGeForce RTX 4070 Ti SuperfitsGeForce RTX 4090fitsGeForce RTX 5070tightGeForce RTX 5080fits

FAQ

Why Q4_K_M?

Q4_K_M is the quant this page recommends. The file is phi-4-Q4_K_M.gguf at 8.43 GiB. Other single-file quants are in the table when a real size was available.

Where do the architecture numbers come from?

Official config.json. Context 16,384. Source: https://huggingface.co/microsoft/phi-4/raw/main/config.json

Does this include a long context?

The fit uses 8,192 tokens, or less when that is all the fast memory can hold. The model file allows up to 16,384 tokens, and KV cache grows as you use more of that.

What is the Ollama command?

ollama run phi4:14b. The library tag is Q4-class or the vendor quant. Compare it with the GGUF linked on this page before you treat the sizes as identical.