Phi · Microsoft
Fine-tuning Phi-4 14B
What does it take to fine-tune Phi-4 14B?
Phi-4 14B needs 28 GB for half-precision LoRA training and 20 GB in four-bit, and 26 GB to serve. It supports supervised fine-tuning, preference tuning (dpo), under the MIT licence. The cheapest qualifying class is RTX 4000 Ada at $0.09 per GPU-hour.
A 14B that benchmarks in the range of 32B models on reasoning-heavy evaluations, under MIT. That combination is the argument for it: you get roughly large-tier behaviour while staying inside the mid-large tier for both memory and hourly rate.
What to know before choosing it
Benchmarking in the range of 32B models on reasoning-heavy evaluations, under MIT, at mid-large memory, is the whole argument. Roughly large-tier behaviour on the tasks Phi is good at, while staying inside a 48GB class and a permissive licence.
The caveat is the Mini's, scaled up: the strength is reasoning and code rather than breadth. On a knowledge-heavy task a Qwen3 or Llama of similar size will often do better, and the gate is the cheapest way to find out which.
Specification
| Parameters | 14B |
|---|---|
| Architecture | Dense |
| Size tier | Mid-large |
| LoRA training memory | 28 GBHalf precision, frozen base |
| QLoRA training memory | 20 GBFour-bit base, higher-precision adapter |
| Serving memory | 26 GBHalf precision, before attention cache |
| Objectives | SFT, DPO |
| Licence | MIT |
| Repository | microsoft/phi-4 |
What it costs to train
Every class with enough memory for four-bit training, cheapest first. Total cost is the rate multiplied by wall time, so the cheapest rate is not always the cheapest run — a faster class that finishes sooner frequently wins.
| GPU class | VRAM | Training / hr | Serving / hr | Fits |
|---|---|---|---|---|
| RTX 4000 Ada | 20 GB | $0.09 | $0.11 | QLoRA only |
| L4 | 24 GB | $0.17 | $0.20 | QLoRA only |
| RTX 3090 | 24 GB | $0.22 | $0.26 | QLoRA only |
| RTX 4090 | 24 GB | $0.38 | $0.43 | QLoRA only |
| A40 | 48 GB | $0.42 | $0.48 | LoRA and QLoRA |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | LoRA and QLoRA |
| A6000 | 48 GB | $0.65 | $0.75 | LoRA and QLoRA |
| L40S | 48 GB | $0.78 | $0.90 | LoRA and QLoRA |
| A100 40 GB | 40 GB | $1.17 | $1.35 | LoRA and QLoRA |
| A100 80 GB | 80 GB | $1.68 | $1.94 | LoRA and QLoRA |
| H100 80 GB | 80 GB | $2.59 | $2.99 | LoRA and QLoRA |
| H200 | 141 GB | $4.55 | $5.25 | LoRA and QLoRA |
Half-precision LoRA needs 28 GB, so the cheapest class for it is A40 at $0.42 per hour. Below that, training has to be quantised.
Serving fits on A40 at $0.48 per hour — before the attention cache, which grows with context length and concurrency.
Good starting point for
- Reasoning tasks that outgrew 8B
- Getting large-model quality at mid-tier cost
- Code generation and review
Preference tuning on this model
Preference tuning holds a frozen reference copy of the model alongside the one being trained, so budget roughly 56 GB rather than 28 GB. That is the thing that catches people out — supervised training on this model fits on a class that preference tuning will overflow.
Other sizes in this family
| Model | Params | QLoRA | Objectives |
|---|---|---|---|
| Phi-4 Mini 3.8B | 3.8B | 8 GB | SFT, DPO |
Comparable sizes elsewhere
Frequently asked questions
How much VRAM does it take to fine-tune Phi-4 14B?
28 GB for half-precision LoRA and 20 GB for four-bit QLoRA. Serving needs 26 GB. Preference tuning roughly doubles the training figure, because a frozen reference model is held alongside the one being trained.
What is the cheapest way to fine-tune Phi-4 14B?
Four-bit QLoRA on RTX 4000 Ada at $0.09 per GPU-hour is the cheapest class that meets the 20 GB threshold. Note that quantised training is slower per step, so a faster class sometimes costs less over the whole run.
Can I download the weights after fine-tuning Phi-4 14B?
Yes. Every finished run exposes its trained weights for download, and publishing to a model hub is a single call with a generated model card recording the base model and version the adapter applies to.
What licence does Phi-4 14B carry?
MIT. The licence follows the fine-tune — a derivative inherits the base model’s terms, and those terms pass to anyone you give the model to.
Last verified 6 August 2026. Memory thresholds are the platform's own admission limits.
Fine-tune Phi-4 14B
Upload a dataset, forecast the run, and see the cost before any compute is leased.