Qwen3 · Alibaba
Fine-tuning Qwen3 4B
What does it take to fine-tune Qwen3 4B?
Qwen3 4B needs 12 GB for half-precision LoRA training and 8 GB in four-bit, and 9 GB to serve. It supports supervised fine-tuning, preference tuning (dpo), reinforcement learning (grpo), under the Apache 2.0 licence. The cheapest qualifying class is RTX 3080 at $0.09 per GPU-hour.
The smallest Qwen3 that reliably handles multi-turn tool calling after tuning. Twelve gigabytes for half-precision LoRA means the entry GPU classes are still in play, so the cost step from 1.7B is smaller than the capability step.
What to know before choosing it
Multi-turn tool calling is the capability that appears here and not below. Holding a tool schema, choosing the right one, formatting the arguments and then handling the result in sequence needs more working capacity than 1.7B has, and the failure below this size is not subtle — arguments come out malformed under pressure.
Twelve gigabytes for half-precision LoRA is the number that makes it attractive: the entry GPU classes still cover it, so the cost step up from 1.7B is much smaller than the capability step.
Specification
| Parameters | 4B |
|---|---|
| Architecture | Dense |
| Size tier | Small |
| LoRA training memory | 12 GBHalf precision, frozen base |
| QLoRA training memory | 8 GBFour-bit base, higher-precision adapter |
| Serving memory | 9 GBHalf precision, before attention cache |
| Objectives | SFT, DPO, GRPO |
| Licence | Apache 2.0 |
| Repository | Qwen/Qwen3-4B |
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 |
|---|---|---|---|---|
| RTX 3080 | 12 GB | $0.09 | $0.11 | LoRA and QLoRA |
| RTX 4000 Ada | 20 GB | $0.09 | $0.11 | LoRA and QLoRA |
| L4 | 24 GB | $0.17 | $0.20 | LoRA and QLoRA |
| RTX 3090 | 24 GB | $0.22 | $0.26 | LoRA and QLoRA |
| RTX 4090 | 24 GB | $0.38 | $0.43 | LoRA and QLoRA |
| A40 | 48 GB | $0.42 | $0.48 | LoRA and QLoRA |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | LoRA and QLoRA |
| A6000 | 48 GB | $0.65 | $0.75 | LoRA and QLoRA |
| L40S | 48 GB | $0.78 | $0.90 | LoRA and QLoRA |
| A100 40 GB | 40 GB | $1.17 | $1.35 | LoRA and QLoRA |
| 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 |
Serving fits on RTX 3080 at $0.11 per hour — before the attention cache, which grows with context length and concurrency.
Good starting point for
- Tool-calling agents
- Summarisation with format constraints
- Cost-sensitive production endpoints
Preference tuning on this model
Preference tuning holds a frozen reference copy of the model alongside the one being trained, so budget roughly 24 GB rather than 12 GB. That is the thing that catches people out — supervised training on this model fits on a class that preference tuning will overflow.
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 8B | 8B | 14 GB | SFT, DPO, GRPO |
| Qwen3 14B | 14B | 20 GB | SFT, DPO, GRPO |
| Qwen3 30B-A3B | 30B | 36 GB | SFT |
| Qwen3 32B | 32B | 48 GB | SFT |
Comparable sizes elsewhere
Frequently asked questions
How much VRAM does it take to fine-tune Qwen3 4B?
12 GB for half-precision LoRA and 8 GB for four-bit QLoRA. Serving needs 9 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 4B?
Four-bit QLoRA on RTX 3080 at $0.09 per GPU-hour is the cheapest class that meets the 8 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 4B?
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 4B 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 4B
Upload a dataset, forecast the run, and see the cost before any compute is leased.