Phi-4.
Phi-4 is Microsoft Research's 14B MIT-licensed text model with a 16K-token context, for memory- and latency-limited uses, focused mainly on English.
1× Nvidia GeForce RTX 3060 · $0.038/hr.
Most-aggressive quantisation we have a working recommendation for. Lower precision = less VRAM = cheaper hardware, at a small accuracy cost.
Cheapest hosted endpoints.
Variants in the Phi family.
Cheapest hardware per quantisation.
Each row is one quantisation tier (the same weights compressed differently). Lower precision → lower VRAM → cheaper hardware, at the cost of small accuracy loss. $/hr is the cheapest rate on offer now. A multi-card setup uses a real instance of that size when one is rented; otherwise it is the single-card rate times the count.
| Quantisation | Cheapest GPU config | Total VRAM | Live $/hr | tokens/sec | |
|---|---|---|---|---|---|
|
FP16
FP16 — half precision (default)
|
48 GB |
$0.26/hr
Vast.ai · on-demand
|
16.0 | Compare → | |
|
FP8
FP8 — 8-bit float (Hopper / Blackwell)
|
24 GB |
$0.14/hr
Vast.ai · on-demand
|
32.0 | Compare → | |
|
INT4
INT4 — 4-bit integer (~4× VRAM saving)
|
12 GB |
$0.038/hr
Nosana · on-demand
|
34.3 | Compare → |
What it costs per month across providers.
Estimate your monthly bill for Phi-4 across every host that publishes per-token pricing. Slide your token volumes; the chart + table re-rank cheapest-first.
Cheapest provider on the left.
Total monthly cost — input + output tokens combined.
Bill breakdown.
Rent the GPU instead of paying per token.
For an open-weights model like Phi-4, you can rent a GPU and serve inference yourself. The math: cheapest GPU rental × 730 hours/month. Power is included in a rental's hourly price.
Assumes the GPU runs 24/7 at ~85% utilisation. If your traffic is bursty, you'll pay less for the API and probably more for the GPU (idle hours still cost rental). The breakeven analysis lives on the Self-host vs API breakeven tool.
About Phi-4.
Phi-4 is a small open model from Microsoft Research under the MIT licence, with 14B parameters in a dense decoder-only design. It takes text, best given in chat format, and has a 16K-token context. Its training blended synthetic data, filtered public web data and acquired academic books, with a focus on reasoning. Microsoft aims it at research and at features with tight memory or latency limits. It is mainly for English, and it is a static model with public data cut off at June 2024.