Can Meta Llama 3.1 8B Instruct run on the 2x RTX 4090?
Yes. At 8K context the best fit is F16, using about 16.6 GB of the 45.6 GB usable on the 2x RTX 4090. Estimated speed ceiling: ~63 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) | 3.9 GB Fits | 4.4 GB Fits | 7.4 GB Fits | 19.4 GB Fits | 335 |
| 3-bit (Q3_K_M) | 4.8 GB Fits | 5.3 GB Fits | 8.3 GB Fits | 20.3 GB Fits | 257 |
| 4-bit (Q4_K_M) • | 5.7 GB Fits | 6.2 GB Fits | 9.2 GB Fits | 21.2 GB Fits | 205 |
| 5-bit (Q5_K_M) • | 6.5 GB Fits | 7.0 GB Fits | 10.0 GB Fits | 22.0 GB Fits | 176 |
| 6-bit (Q6_K) • | 7.3 GB Fits | 7.8 GB Fits | 10.8 GB Fits | 22.8 GB Fits | 153 |
| 8-bit (Q8_0) • | 9.1 GB Fits | 9.6 GB Fits | 12.6 GB Fits | 24.6 GB Fits | 118 |
| 16-bit (F16/BF16) • | 16.1 GB Fits | 16.6 GB Fits | 19.6 GB Fits | 31.6 GB Fits | 63 |
• real GGUF file size from Saxo/Linkbricks-Horizon-AI-Korean-llama3.1-sft-rlhf-dpo-8B. Speed is a bandwidth ceiling (1008 GB/s), not a benchmark. Methodology.
More
Meta Llama 3.1 8B Instruct on every GPU and Mac · Every model that runs on the 2x RTX 4090 · Try other settings in the calculator
Data: Hugging Face Hub (NousResearch/Meta-Llama-3.1-8B-Instruct), NVIDIA specs. Model data last checked 2026-11-17; refreshed hourly.