Microsoft · 14B parameters · 16.4K context · commercial OK
Local (your GPU) $0 · ~13 GB VRAM | Rent a GPU from $0.12/hr | Your API key $0.11/1M |
Best next step
Already have 13 GB of VRAM? Run Q4_K_M locally for $0 compute. Open the local setup guide →
No matching GPU? For a 4,000-token prompt + 2,000-token answer, an API is about $0.0006 — no GPU setup. Open the API option →
Decision basis: 4,000 input + 2,000 output tokens, prices checked 2026-07-27. API estimates never beat a real price here; referral status never changes the winner.
Novita AI offers model APIs and GPU instances. We do not include its prices in Spanvero's ranking yet, so compare Novita's current price and fit yourself.
Check Novita AI — referral link. You use your own account and pay Novita's normal price; Novita may pay Spanvero a commission.
Watch this price — free
Get an email when Phi-4's tracked price drops — checked against the append-only daily record. No account needed.
Free accounts watch one price. Exact multi-price targets remain available to existing Pro accounts; the old subscription is no longer sold — see current products.
Microsoft’s 14B reasoning-focused model, MIT-licensed. Strong logic/math for its size.
$0.105/1M blended input + output
Published price logged daily since June 23, 2026. See the full price record.
At $0.11/1M tokens, the cheapest API for Phi-4 stays cheaper than a used RTX 3090 at any daily volume — electricity alone costs more per token than the API does.
Local is never $0 — every figure here includes electricity and hardware amortization. Speeds come from published 7-8B Q4 benchmarks (your model, quant, and context will differ); break-even volumes are rounded to 2 significant figures.
| Parameters | 14B |
|---|---|
| Context window | 16.4K tokens |
| Recommended quant | Q4_K_M |
| VRAM requirement | ~13 GB (at Q4_K_M, 16.4K context) |
| Download size | ~9.1 GB |
| License | Commercial use OK |
Open the free Spanvero advisor → for the live, interactive math for your exact workload and hardware.
Three ways: $0 on your own machine if you have about 13 GB of VRAM (at Q4_K_M); a rented GPU from $0.12/hr (2× NVIDIA RTX 3060 12GB at the direct Vast.ai price); or about $0.11 per 1M blended tokens via your own API key (real price as of 2026-07-27). Spanvero adds $0 markup on every path.
Yes — Phi-4 runs locally for $0 in compute if your GPU or Apple Silicon Mac has about 13 GB of VRAM/unified memory (at Q4_K_M with 16.4K context). Point LM Studio, Ollama, or llama.cpp at it; after the download it runs offline and free.
Any GPU or Apple Silicon Mac with about 13 GB of VRAM/unified memory handles Phi-4 at Q4_K_M. Of the consumer hardware Spanvero tracks, the smallest that fits is NVIDIA RTX 4060 Ti 16GB.
Browse: How to run Phi-4 locally · More Microsoft (Phi) models · All models · Compare
Spanvero · What's new · Prices as of 2026-07-27. We're an honest advisor — $0 markup, your own accounts, we never resell compute. © 2026 Cynosure LLC.
A short email of real AI price moves, straight from the daily log — no hype. We're collecting the list now; the first issue goes out when it opens. Unsubscribe with one click.