Qwen3 · Alibaba
Fine-tuning Qwen3 30B-A3B
What does it take to fine-tune Qwen3 30B-A3B?
Qwen3 30B-A3B needs 64 GB for half-precision LoRA training and 36 GB in four-bit, and 48 GB to serve. It supports supervised fine-tuning, under the Apache 2.0 licence. The cheapest qualifying class is A40 at $0.42 per GPU-hour.
Thirty billion total parameters, three billion active per token. The trade is explicit: you pay 30B memory during training and serving, but roughly 3B compute per token at inference. That makes it interesting when throughput matters more than memory cost, and uninteresting when the reverse is true.
What to know before choosing it
Thirty billion resident, three billion active. That ratio is the entire decision: you pay large-tier memory and mid-tier compute per token. It is a good trade when an endpoint is busy and a poor one when it is idle, because memory is charged for whether or not tokens are flowing.
Fine-tuning it is also harder than the dense models around it. Routing between experts can become unbalanced during training, concentrating learning on a subset, and the symptom — quality plateauing early while loss still falls — looks like several other problems.
Specification
| Parameters | 30B3B active per token |
|---|---|
| Architecture | Mixture of experts |
| Size tier | Extra large |
| LoRA training memory | 64 GBHalf precision, frozen base |
| QLoRA training memory | 36 GBFour-bit base, higher-precision adapter |
| Serving memory | 48 GBHalf precision, before attention cache |
| Objectives | SFT |
| Licence | Apache 2.0 |
| Repository | Qwen/Qwen3-30B-A3B |
What it costs to train
Every class with enough memory for four-bit training, cheapest first. Total cost is the rate multiplied by wall time, so the cheapest rate is not always the cheapest run — a faster class that finishes sooner frequently wins.
| GPU class | VRAM | Training / hr | Serving / hr | Fits |
|---|---|---|---|---|
| A40 | 48 GB | $0.42 | $0.48 | QLoRA only |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | QLoRA only |
| A6000 | 48 GB | $0.65 | $0.75 | QLoRA only |
| L40S | 48 GB | $0.78 | $0.90 | QLoRA only |
| A100 40 GB | 40 GB | $1.17 | $1.35 | QLoRA only |
| A100 80 GB | 80 GB | $1.68 | $1.94 | LoRA and QLoRA |
| H100 80 GB | 80 GB | $2.59 | $2.99 | LoRA and QLoRA |
| H200 | 141 GB | $4.55 | $5.25 | LoRA and QLoRA |
Half-precision LoRA needs 64 GB, so the cheapest class for it is A100 80 GB at $1.68 per hour. Below that, training has to be quantised.
Serving fits on A40 at $0.48 per hour — before the attention cache, which grows with context length and concurrency.
Good starting point for
- High-throughput serving where memory is cheaper than compute
- Tasks benefiting from broad knowledge at low per-token cost
This is a mixture-of-experts model
30B total parameters with 3B active per token. Memory scales with the total; inference compute scales with the active count. That makes it attractive when memory is cheaper than compute for your workload, and unattractive when the reverse holds. Quoting only one of the two numbers is how these models get misrepresented in both directions.
Other sizes in this family
| Model | Params | QLoRA | Objectives |
|---|---|---|---|
| Qwen3 0.6B | 600M | 3 GB | SFT, DPO, GRPO |
| Qwen3 1.7B | 1.7B | 4 GB | SFT, DPO, GRPO |
| Qwen3 4B | 4B | 8 GB | SFT, DPO, GRPO |
| Qwen3 8B | 8B | 14 GB | SFT, DPO, GRPO |
| Qwen3 14B | 14B | 20 GB | SFT, DPO, GRPO |
| Qwen3 32B | 32B | 48 GB | SFT |
Comparable sizes elsewhere
Frequently asked questions
How much VRAM does it take to fine-tune Qwen3 30B-A3B?
64 GB for half-precision LoRA and 36 GB for four-bit QLoRA. Serving needs 48 GB. Preference tuning roughly doubles the training figure, because a frozen reference model is held alongside the one being trained.
What is the cheapest way to fine-tune Qwen3 30B-A3B?
Four-bit QLoRA on A40 at $0.42 per GPU-hour is the cheapest class that meets the 36 GB threshold. Note that quantised training is slower per step, so a faster class sometimes costs less over the whole run.
Can I download the weights after fine-tuning Qwen3 30B-A3B?
Yes. Every finished run exposes its trained weights for download, and publishing to a model hub is a single call with a generated model card recording the base model and version the adapter applies to.
What licence does Qwen3 30B-A3B carry?
Apache 2.0. The licence follows the fine-tune — a derivative inherits the base model’s terms, and those terms pass to anyone you give the model to.
Last verified 6 August 2026. Memory thresholds are the platform's own admission limits.
Fine-tune Qwen3 30B-A3B
Upload a dataset, forecast the run, and see the cost before any compute is leased.