Phi-4 14B
Microsoft
Trained largely on curated synthetic data. Punches far above its size on reasoning and maths; the short context limits what you can do with it.
Good fit
- Weights
- 15 GB
- KV cache
- 1.6 GB
- Overhead
- 710 MB
- 17 GB of 23 GB usable VRAM.
- Chosen as the best quality that still fits a 16,384-token context (18 GB at that length).
- Exceptional STEM reasoning for 14B
- MIT licensed
- Fits 12 GB at Q4_K_M
- 16k context
- Narrower world knowledge than web-trained peers
- Weaker multilingual coverage
phi4:14bRunning Phi-4 14B on GeForce RTX 4090
| Quantization | Quality | Weights | KV cache | Total | ~tok/s | Fit |
|---|---|---|---|---|---|---|
| F16 | lossless | 27 GB | 1.6 GB | 30 GB | 27 | Runs with CPU offload |
| Q8_0Pick | near-lossless | 15 GB | 1.6 GB | 17 GB | 50 | Good fit |
| Q6_K | near-lossless | 11 GB | 1.6 GB | 14 GB | 65 | Excellent fit |
| Q5_K_M | high | 9.7 GB | 1.6 GB | 12 GB | 75 | Excellent fit |
| Q4_K_M | balanced | 8.3 GB | 1.6 GB | 11 GB | 88 | Excellent fit |
| Q3_K_M | degraded | 6.7 GB | 1.6 GB | 9.0 GB | 108 | Excellent fit |
KV cache is sized at 8,192 tokens. Longer contexts cost proportionally more — the recommendation above reserves room for a working context.
You can fine-tune this here
- Base weights
- 7.7 GB
- Optimizer
- 977 MB
- Activations
- 1.3 GB
- Peak
- 12 GB
- 12 GB peak against 23 GB usable — room to raise batch size or sequence length.
- Base weights
- 27 GB
- Optimizer
- 977 MB
- Activations
- 1.3 GB
- Peak
- 31 GB
- Needs 31 GB — switch to QLoRA to cut the weight footprint.
Where this model runs
VRAM 32 GB · Q8_0 · 17 GB
VRAM 24 GB · Q8_0 · 17 GB
VRAM 16 GB · Q5_K_M · 12 GB
VRAM 16 GB · Q5_K_M · 12 GB
VRAM 24 GB · Q8_0 · 17 GB
VRAM 12 GB · Q4_K_M · 11 GB
VRAM 12 GB · Q4_K_M · 11 GB
VRAM 16 GB · Q5_K_M · 12 GB
VRAM 48 GB · F16 · 30 GB
VRAM 80 GB · F16 · 30 GB
Unified 128 GB · F16 · 30 GB
Unified 48 GB · Q8_0 · 17 GB
Unified 24 GB · Q5_K_M · 12 GB
Unified 192 GB · F16 · 30 GB
Unified 16 GB · Q4_K_M · 11 GB
VRAM 24 GB · Q8_0 · 17 GB
VRAM 16 GB · Q5_K_M · 12 GB
VRAM 0 MB
VRAM 0 MB
84.8%
Phi-4 technical report
Reported by the model's author. ModelLM has not run these benchmarks and does not treat them as verified.
Qwen2.5 14B Instruct
14.8B · Apache 2.0
The sweet spot for 24 GB cards. Meaningfully stronger reasoning than 7B while still fine-tunable locally with QLoRA.
Qwen3 14B
14.8B · Apache 2.0
The Qwen3 mid-size. Long context and a reasoning mode inside a footprint a 24 GB card handles comfortably.
Mistral Nemo 12B Instruct
12.2B · Apache 2.0
A 12B with a 128k context and the Tekken tokenizer, which compresses non-English text far better than Llama’s.
Gemma 2 9B Instruct
9.2B · Gemma Terms of Use
Unusually good at natural, well-structured prose for its size. The short context is the catch.
Catalogue figures come from each model’s published card. ModelLM has not independently measured them.