Qwen3 30B-A3B — Specifiche
| Sviluppatore | Alibaba |
|---|---|
| Tipo | LLM (MoE) |
| Modalità | Testo → Testo |
| Parametri | 30 miliardi totali / 3 miliardi attivi (MoE) |
| Finestra contestuale | 128K |
| Output massimo | — |
| Licenza | Apache 2.0 (open) |
| Pesi aperti | Sì |
| Pubblicato | 2025 |
| Prezzo dell’input | 0,12 $ /1M |
| Prezzo dell’output | $0,50 /1 milione |
| Provider API | Alibaba, OpenRouter, Ollama |
Esegui localmente
| VRAM (4-bit) | ~18 GB |
|---|---|
| GPU minima richiesta | RTX 4090 24 GB (Q4) — veloce, 3 miliardi attivi |
What is Qwen3 30B-A3B?
Qwen3 30B-A3B is a mixture-of-experts with 30B total parameters but only about 3B active
per token — near-32B quality at a fraction of the compute, and very fast locally. Apache 2.0,
128K context, roughly 18 GB of VRAM at 4-bit on an RTX 4090 24GB.
The speed is the reason to pick it. Because only ~3B parameters are active on each token,
generation runs several times faster than a dense 32B on the same card, while memory use is
only modestly higher. That combination matters specifically for local interactive work —
coding assistants, chat interfaces, anything where you are watching tokens appear — where a
dense 32B on consumer hardware is often just slow enough to be irritating. The trade is that
you pay for 30B of memory to get 3B of speed, so on a 24 GB card the dense Qwen3 32B remains
the better choice if raw quality per gigabyte matters more than latency. Choose 30B-A3B when
responsiveness is the feature.
Qwen3 30B-A3B pricing: API cost per 1M tokens
| Input (per ogni milione di token) | $0.120 |
|---|---|
| Output (per ogni milione di token) | $0.500 |
| Output/input ratio | 4.2× |
| Blended (4:1 in:out) | $0.196 per 1M tokens |
What Qwen3 30B-A3B costs per month
Real monthly spend at a 4:1 input-to-output mix — the ratio a typical chat or RAG workload actually produces.
| Carico di lavoro | Token/mese | Cost / month |
|---|---|---|
| Progetto secondario | 1 milione in ingresso / 0,25 milioni in uscita | $0.25 |
| Piccolo team | 20 milioni in ingresso / 5 milioni in uscita | $4.90 |
| Produzione | 200 milioni in ingresso / 50 milioni in uscita | $49 |
Run your own numbers in the Calcolatore dei costi delle API per l'IA.
Cheaper alternatives to Qwen3 30B-A3B
| Modello | Costo combinato ($/1 milione) | You save |
|---|---|---|
| Mistral 7B aperta | $0.0220 | 89% cheaper |
| Llama 3.1 8B aperta | $0.0220 | 89% cheaper |
| Mistral NeMo 12B aperta | $0.0240 | 88% cheaper |
Self-host or pay the API?
Qwen3 30B-A3B is open-weight, so you can run it yourself. It needs ~18 GB of VRAM at 4-bit (RTX 4090 24GB (Q4) — fast, 3B active). Self-hosting only beats the API once your volume is high enough to keep that hardware busy — the calcolatore self-hosting vs API works out the break-even point for your token volume.
Domande frequenti
How much does Qwen3 30B-A3B cost per 1M tokens?
Qwen3 30B-A3B costs $0.120 per 1M input tokens and $0.500 per 1M output tokens. At a typical 4:1 input-to-output mix that blends to about $0.196 per 1M tokens.
How much does Qwen3 30B-A3B cost per month?
A small-team workload of 20M input and 5M output tokens a month costs about $4.90 on Qwen3 30B-A3B. A side project (1M in / 0.25M out) costs roughly $0.25.
What is a cheaper alternative to Qwen3 30B-A3B?
Mistral 7B is the strongest cheaper option in our database at $0.0220 per 1M blended — about 89% less than Qwen3 30B-A3B. It is also open-weight, so self-hosting is an option.
Can I run Qwen3 30B-A3B locally?
Yes. Qwen3 30B-A3B is open-weight and needs about ~18 GB of VRAM at 4-bit quantisation (RTX 4090 24GB (Q4) — fast, 3B active).
Why does Qwen3 30B-A3B 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 30B-A3B charges 4.2× more for output, which is why prompt-heavy workloads are far cheaper to run than generation-heavy ones.
Prices are the published list rates for the model's primary API and are reviewed as providers change them. Volume, batch and cached-input discounts are not included. Compare every model side by side in the Database di modelli di intelligenza artificiale o il Classifica LLM.
