Skip to content

VRAM reference

Llama 3.2 3B on an RTX 4090

Yes, you can fine-tune Llama 3.2 3B on one RTX 4090 (24 GB) with QLoRA or LoRA. 4-bit inference: yes, 1.8 GB of weights, room for 128K tokens of context or 42 concurrent 4K-token sequences.

By Will Schulz · Checked 15 November 2026 · Prices as of November 2026

Verdict by method

Fine-tuningMemory at 2K tokensOne RTX 4090?Longest sequence (batch 1)Largest batch at 2KLargest LoRA rank
QLoRA (4-bit base)5.3 GBYes8K tokens7512
LoRA (16-bit base)9.5 GBYes8K tokens5512
Full fine-tuning54 GBNo——n/a
InferenceMemory at 8KOne RTX 4090?Longest contextConcurrent 4K sequences
16-bit (BF16/FP16)7.4 GBYes128K tokens32
8-bit4.2 GBYes128K tokens39
4-bit2.7 GBYes128K tokens42
  • QLoRA: yes, about 5.3 GB at 2K tokens, batch 1, rank 16; up to 8K-token sequences at batch 1 or batch 7 at 2K, LoRA rank up to 512.
  • LoRA: yes, about 9.5 GB at 2K tokens, batch 1, rank 16; up to 8K-token sequences at batch 1 or batch 5 at 2K, LoRA rank up to 512.
  • Full fine-tuning: no, it needs about 54 GB at 2K tokens, more than the 24 GB available.

Where the memory goes

At 2,048 tokens per sequence, batch 1, LoRA rank 16, gradient checkpointing on
MethodWeightsTraining stateActivationsTotalOn one RTX 409080 GB GPUs needed
QLoRA (4-bit base)2.2 GB0.4 GB2.7 GB5.3 GBFits1
LoRA (16-bit base)6.4 GB0.4 GB2.7 GB9.5 GBFits1
Full fine-tuning6.4 GB45 GB2.7 GB54 GBDoesn't fit1
One sequence, fp16 KV cache
PrecisionWeightsKV cache at 8KTotalAt 128K contextOn one RTX 4090
16-bit (BF16/FP16)6.4 GB0.9 GB7.4 GB21 GBFits
8-bit3.2 GB0.9 GB4.2 GB18 GBFits
4-bit1.8 GB0.9 GB2.7 GB17 GBFits

The cheapest way to run it

Renting an RTX 4090 costs $0.89 an hour on Runpod Secure Cloud ($0.34 on Runpod Community Cloud). LoRA on it processes about 5,142 tokens a second by our estimate, so 10 million training tokens take 32 minutes and cost about $0.48.

Job (default settings)MemoryCheapest rental that fitsPer hour
QLoRA fine-tuning5.3 GBRTX 3090$0.50
LoRA fine-tuning9.5 GBRTX 3090$0.50
4-bit inference2.7 GBRTX 3090$0.50
16-bit inference7.4 GBRTX 3090$0.50

On-demand rates seen on 15 November 2026: Runpod Secure Cloud, Lambda for the A100 40GB. Consumer cards count as single GPUs only; data-centre GPUs up to eight per node.

Try other settings

JobMemoryRTX 4090 (24 GB)Smallest single GPU
QLoRA (4-bit base)2K tokens × batch 1, rank 165.3 GBFitsRTX 3090 (24 GB)
LoRA (16-bit base)2K tokens × batch 1, rank 169.5 GBFitsRTX 3090 (24 GB)
Full fine-tuning2K tokens × batch 154 GB3 GPUsA100 80GB (80 GB)
Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence7.4 GBFitsRTX 3090 (24 GB)
Inference, 8-bit8K-token context × 1 sequence4.2 GBFitsRTX 3090 (24 GB)
Inference, 4-bit8K-token context × 1 sequence2.7 GBFitsRTX 3090 (24 GB)

A job fits when the estimate is within 90% of the card's memory. Training assumes gradient checkpointing and FlashAttention; multi-GPU counts assume the memory splits evenly (FSDP, ZeRO-3 or tensor parallelism).

How we estimate it

Every figure comes from the model's published architecture (layers, hidden size, attention and KV heads, vocabulary) rather than a rule of thumb, using the same functions as the calculator above. Sizes are in GB (10⁹ bytes).

  • Full fine-tuning: 16 bytes per parameter for mixed-precision Adam (bf16 weights and gradients, fp32 master weights and two moments), as in the ZeRO paper.
  • LoRA: frozen bf16 weights (2 bytes per parameter) plus 16 bytes per adapter parameter, rank 16 on every linear layer. For mixture-of-experts models, adapters go on attention only and the experts stay frozen.
  • QLoRA: NF4 weights with double quantisation (4.127 bits per parameter); embeddings and the output head stay in bf16, as bitsandbytes leaves them.
  • Activations: gradient checkpointing and FlashAttention. Each layer's bf16 input is kept, one layer is recomputed at a time, and the output logits take 8 bytes per vocabulary entry per token (fp32 values and gradients). Chunked cross-entropy, as in Unsloth or Liger kernels, cuts that last term.
  • Inference: 2, 1 or about 0.56 bytes per weight (4-bit with group-wise scales), plus an fp16 KV cache of 2 × layers × KV heads × head dim × 2 bytes per token; sliding-window layers only cache their window.
  • Fits means the estimate is within 90% of the card's listed memory; the rest covers the CUDA context, kernels and fragmentation. Multi-GPU counts assume memory splits evenly.

Speed estimates use RTX 4090's dense BF16 peak of 165.2 TFLOPS, about 4 FLOPs per active parameter per token for LoRA and 40% utilisation; see fine-tuning and training costs for the method.

Frequently asked questions

Can I fine-tune Llama 3.2 3B on an RTX 4090?
QLoRA: yes, about 5.3 GB at 2K tokens, batch 1, rank 16; up to 8K-token sequences at batch 1 or batch 7 at 2K, LoRA rank up to 512. LoRA: yes, about 9.5 GB at 2K tokens, batch 1, rank 16; up to 8K-token sequences at batch 1 or batch 5 at 2K, LoRA rank up to 512. Full fine-tuning: no, it needs about 54 GB at 2K tokens, more than the 24 GB available.
Can I run Llama 3.2 3B on an RTX 4090?
16-bit (BF16/FP16) inference: yes, 6.4 GB of weights, room for 128K tokens of context or 32 concurrent 4K-token sequences. 8-bit inference: yes, 3.2 GB of weights, room for 128K tokens of context or 39 concurrent 4K-token sequences. 4-bit inference: yes, 1.8 GB of weights, room for 128K tokens of context or 42 concurrent 4K-token sequences.
What does fine-tuning Llama 3.2 3B on an RTX 4090 cost?
With LoRA at our planning rate of 40% utilisation, about 32 minutes for 10 million training tokens, around $0.48 at $0.89 an hour (Runpod Secure Cloud, November 2026).

Sources

  1. [1]Hugging Face. meta-llama/Llama-3.2-3B-Instruct config.json
  2. [2]Meta. meta-llama/Llama-3.2-3B-Instruct model card
  3. [3]Meta. The Llama 3 Herd of Models (arXiv:2407.21783)
  4. [4]NVIDIA. RTX 4090 specifications
  5. [5]Runpod. GPU pricing (updated 12 November 2026)
  6. [6]Lambda. GPU cloud pricing
  7. [7]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  8. [8]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  9. [9]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  10. [10]Korthikanti et al., 2022. Reducing Activation Recomputation in Large Transformer Models (arXiv:2205.05198)

Measure before you rent the GPU.

Prompt tuning and model search in Kojen run on hosted models against your own test cases, so you know whether fine-tuning Llama 3.2 3B is worth the GPU time.

Start training free