Qwen3 32B — Specifiche
Compiled by Mustafa Ihsan from the vendor’s published documentation · Last updated
| Sviluppatore | Alibaba |
|---|---|
| Tipo | LLM (densa) |
| Modalità | Testo → Testo |
| Parametri | 32 miliardi |
| Finestra contestuale | 128K |
| Licenza | Apache 2.0 (open) |
| Pesi aperti | Sì |
| Pubblicato | 2025 |
| Prezzo dell’input | $0.08 /1M |
| Prezzo dell’output | $0.28 /1M |
| Provider API | Alibaba, OpenRouter, Ollama |
Esegui localmente
| VRAM (4-bit) | ~20 GB |
|---|---|
| GPU minima | RTX 4090 24 GB (Q4) |
What is Qwen3 32B?
Qwen3 32B is Alibaba’s largest dense Qwen3 — strong general reasoning and coding under a
fully permissive Apache 2.0 licence, with a 128K context. It runs comfortably at 4-bit on a
24 GB consumer GPU, needing about 20 GB, and costs $0.08 in / $0.28 out per million tokens
hosted.
This is the model that makes an RTX 4090 or 3090 worth owning for inference. At 20 GB it
leaves just enough headroom on a 24 GB card for a real context window, and being dense, its
quality per gigabyte is the best available in that bracket — mixture-of-experts models of
similar capability need far more memory resident even though they compute less per token. For
local coding assistants and private document work on a single consumer card, it is close to
the practical maximum. Apache 2.0 removes any licensing review, and the hosted price is low
enough to prototype against before buying hardware. Check your intended quantisation against
il Calcolatore VRAM — the margin on a
24 GB card is real but not generous.
Qwen3 32B pricing: API cost per 1M tokens
| Input (per ogni milione di token) | $0.0800 |
|---|---|
| Output (per ogni milione di token) | $0.280 |
| Rapporto output/input | 3.5× |
| Combinato (4:1 in:out) | $0.120 per 1 milione di token |
What Qwen3 32B costs per month
Spesa mensile reale con un mix input-to-output 4:1 — il rapporto effettivamente prodotto da un tipico carico di lavoro basato su chat o RAG.
| Carico di lavoro | Token/mese | Costo/mese |
|---|---|---|
| Progetto secondario | 1 milione in ingresso / 0,25 milioni in uscita | $0.15 |
| Piccolo team | 20 milioni in ingresso / 5 milioni in uscita | $3.00 |
| Produzione | 200 milioni in ingresso / 50 milioni in uscita | $30 |
Calcola i tuoi numeri personalizzati nel Calcolatore dei costi delle API per l'IA.
Cheaper alternatives to Qwen3 32B
| Modello | Costo combinato ($/1 milione) | Risparmi |
|---|---|---|
| Gemma 3 27B aperta | $0.0960 | 20% cheaper |
Eseguirlo in autonomia o pagare l’API?
Qwen3 32B is open-weight, so you can run it yourself. It needs ~20 GB of VRAM at 4-bit (RTX 4090 24GB (Q4)). Self-hosting only beats the API once your volume is high enough to keep that hardware busy — the calcolatore self-hosting vs API calcola il punto di pareggio in base al tuo volume di token.
Domande frequenti
How much does Qwen3 32B cost per 1M tokens?
Qwen3 32B costs $0.0800 per 1M input tokens and $0.280 per 1M output tokens. At a typical 4:1 input-to-output mix that blends to about $0.120 per 1M tokens.
How much does Qwen3 32B cost per month?
A small-team workload of 20M input and 5M output tokens a month costs about $3.00 on Qwen3 32B. A side project (1M in / 0.25M out) costs roughly $0.15.
What is a cheaper alternative to Qwen3 32B?
Gemma 3 27B is the strongest cheaper option in our database at $0.0960 per 1M blended — about 20% less than Qwen3 32B. It is also open-weight, so self-hosting is an option.
Can I run Qwen3 32B locally?
Yes. Qwen3 32B is open-weight and needs about ~20 GB of VRAM at 4-bit quantisation (RTX 4090 24GB (Q4)).
Why does Qwen3 32B 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. Qwen3 32B charges 3.5× more for output, which is why prompt-heavy workloads are far cheaper to run than generation-heavy ones.
I prezzi indicati sono le tariffe ufficiali pubblicate per l'API principale del modello e vengono aggiornati man mano che i fornitori li modificano. Sconti per volumi elevati, elaborazione batch e input memorizzati nella cache non sono inclusi. Confronta tutti i modelli fianco a fianco nel Database di modelli IA o il Classifica LLM.