Phi-4 — Spezifikationen
Compiled by Mustafa Ihsan from the vendor’s published documentation · Last updated
| Entwickler | Microsoft |
|---|---|
| Typ | LLM (dicht) |
| Modality | Text → Text |
| Parameter | 14B |
| Kontextfenster | 16K |
| Lizenz | MIT (offen) |
| Offene Gewichte | Ja |
| Veröffentlicht | 2025 |
| Eingabepreis | $0.07 /1M |
| Ausgabepreis | $0.14 /1M |
| API-Anbieter | Azure, OpenRouter, Ollama |
Lokal ausführen
| VRAM (4-Bit) | ~9 GB |
|---|---|
| Mindest-GPU | RTX 4070 12 GB / RTX 3060 12 GB |
What is Phi-4?
Phi-4 is Microsoft’s compact 14B reasoning model, MIT-licensed, which punches well above
its size on mathematics and logic. It needs about 9 GB of VRAM at 4-bit, running comfortably
on an RTX 4070 or RTX 3060 12GB, and costs $0.07 in / $0.14 out per million tokens
hosted.
The Phi line’s whole thesis is that curated, textbook-quality training data beats raw
scale for reasoning tasks, and Phi-4 is the clearest evidence for it — a 14B model competing
on maths and logic benchmarks with models several times larger. The cost of that focus is the
16K context window, by far the narrowest in this database and a hard limit for any workload
involving documents, long conversations or retrieval. Read it as a specialist: excellent for
structured reasoning over short inputs — maths tutoring, logic and code puzzles, deterministic
extraction from small payloads — and the wrong tool the moment your prompt grows. If you need
Phi-4’s reasoning with room to work, a 128K-context model in the same hardware bracket such
as Qwen3 14B is the better trade.
Phi-4 pricing: API cost per 1M tokens
| Eingabe (pro 1 Mio. Token) | $0.0700 |
|---|---|
| Ausgabe (pro 1 Mio. Token) | $0.140 |
| Verhältnis Output/Input | 2× |
| Gemischt (4:1 Input:Output) | $0.0840 pro 1 Mio. Tokens |
What Phi-4 costs per month
Tatsächliche monatliche Ausgaben bei einem 4:1-Input-zu-Output-Mix – dem Verhältnis, das typische Chat- oder RAG-Arbeitslasten tatsächlich erzeugen.
| Workload | Tokens/Monat | Kosten pro Monat |
|---|---|---|
| Nebenprojekt | 1 Mio. Eingabe / 0,25 Mio. Ausgabe | $0.11 |
| Kleines Team | 20 Mio. Eingabe / 5 Mio. Ausgabe | $2.10 |
| Produktion | 200 Mio. Eingabe / 50 Mio. Ausgabe | $21 |
Stellen Sie Ihre eigenen Berechnungen im KI-API-Kostenrechner.
Cheaper alternatives to Phi-4
| Modell | Gewichteter Preis pro Million US-Dollar | Sie sparen |
|---|---|---|
| Mistral 7B offenere | $0.0220 | 74% cheaper |
| Llama 3.1 8B offenere | $0.0220 | 74% cheaper |
| Mistral NeMo 12B offenere | $0.0240 | 71 % günstiger |
Selbst hosten oder die API nutzen?
Phi-4 is open-weight, so you can run it yourself. It needs ~9 GB of VRAM at 4-bit (RTX 4070 12GB / RTX 3060 12GB). Self-hosting only beats the API once your volume is high enough to keep that hardware busy — the Selbsthosting vs. API-Rechner berechnet den Break-even-Punkt für Ihr Token-Volumen.
Häufig gestellte Fragen
How much does Phi-4 cost per 1M tokens?
Phi-4 costs $0.0700 per 1M input tokens and $0.140 per 1M output tokens. At a typical 4:1 input-to-output mix that blends to about $0.0840 per 1M tokens.
How much does Phi-4 cost per month?
A small-team workload of 20M input and 5M output tokens a month costs about $2.10 on Phi-4. A side project (1M in / 0.25M out) costs roughly $0.11.
What is a cheaper alternative to Phi-4?
Mistral 7B is the strongest cheaper option in our database at $0.0220 per 1M blended — about 74% less than Phi-4. It is also open-weight, so self-hosting is an option.
Can I run Phi-4 locally?
Yes. Phi-4 is open-weight and needs about ~9 GB of VRAM at 4-bit quantisation (RTX 4070 12GB / RTX 3060 12GB).
Why does Phi-4 charge more for output than input?
Output tokens are generated one at a time and cannot be batched the way a prompt can, so they cost the provider more to serve. Phi-4 charges 2× more for output, which is why prompt-heavy workloads are far cheaper to run than generation-heavy ones.
Die Preise entsprechen den offiziell veröffentlichten Listenpreisen für die primäre API des jeweiligen Modells und werden regelmäßig aktualisiert, sobald Anbieter diese ändern. Volumen-, Batch- und Cached-Input-Rabatte sind nicht enthalten. Vergleichen Sie alle Modelle nebeneinander im Datenbank für KI-Modelle oder das LLM-Leaderboard.