Qwen3 · Alibaba
Fine-tuning Qwen3 8B
What does it take to fine-tune Qwen3 8B?
Qwen3 8B needs 18 GB for half-precision LoRA training and 14 GB in four-bit, and 16 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 4000 Ada at $0.09 per GPU-hour.
The Qwen team's own framing is that the 8B base performs around the level of the previous generation's 14B. It is Apache-2.0, supports every objective offered here, and fits half-precision LoRA on a single 24GB card — which is why it is the most common starting point on the platform for anyone without a specific reason to use Llama.
What to know before choosing it
The comparison the Qwen team draws — the 8B base performing around the level of the previous generation's 14B — is the practical reason this shows up as the default so often. You get mid-large behaviour at mid-tier memory, under a licence with no acceptable-use addendum, no naming requirement and no revenue threshold.
The other reason is objective coverage. It is one of the few models here supporting supervised, preference and reinforcement learning, so a project can start supervised and later add a reward function without changing base model — and changing base model mid-project costs more than people expect.
Specification
| Parameters | 8B |
|---|---|
| Architecture | Dense |
| Size tier | Mid |
| LoRA training memory | 18 GBHalf precision, frozen base |
| QLoRA training memory | 14 GBFour-bit base, higher-precision adapter |
| Serving memory | 16 GBHalf precision, before attention cache |
| Objectives | SFT, DPO, GRPO |
| Licence | Apache 2.0 |
| Repository | Qwen/Qwen3-8B |
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 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 4000 Ada at $0.11 per hour — before the attention cache, which grows with context length and concurrency.
Good starting point for
- General-purpose production fine-tunes
- Reasoning tasks with reinforcement learning
- Work where licence terms have to be unrestrictive
Preference tuning on this model
Preference tuning holds a frozen reference copy of the model alongside the one being trained, so budget roughly 36 GB rather than 18 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 4B | 4B | 8 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 8B?
18 GB for half-precision LoRA and 14 GB for four-bit QLoRA. Serving needs 16 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 8B?
Four-bit QLoRA on RTX 4000 Ada at $0.09 per GPU-hour is the cheapest class that meets the 14 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 8B?
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 8B 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 8B
Upload a dataset, forecast the run, and see the cost before any compute is leased.