Can Llama 4 Maverick 17B 128E Instruct run on the RTX 3090 Ti?
Not fully in VRAM. Even at 2-bit with 8K context, Llama 4 Maverick 17B 128E Instruct needs about 138 GB, more than the 22.8 GB usable on the RTX 3090 Ti. It can still run with offload (slower): Q2_K with part of the model in system RAM, about 19–35 tokens/sec (estimate). Details and command.
Runs with offload (slower)
Runs with offload 2-bit (Q2_K) at 8K context: all attention layers plus the experts of 2 of 24 MoE layers on the RTX 3090 Ti (about 17.5 GB), and the experts of the other 22 layers in system RAM (about 120 GB). Needs a PC with at least 128 GB of system RAM. Estimated speed: 19–35 tokens/sec (usable for chat), versus a bandwidth ceiling of ~154 tok/s if it fit fully in VRAM.
llama-server -hf unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF:Q2_K -c 8192 -ngl 999 --n-cpu-moe 46 --no-mmproj
Estimate, not a benchmark. The command downloads the GGUF from Hugging Face. The speed range spans dual-channel DDR4-3200 to DDR5-6000 system RAM; your CPU, RAM speed and llama.cpp version move it. If llama.cpp runs out of GPU memory, raise --n-cpu-moe by 1-2; if VRAM is left over, lower it for more speed. Recent llama.cpp builds can also pick the split automatically. How we estimate offload.
+ 1.6 GB with image input. The figures on this page are for text only. With image input enabled, the vision projector (mmproj) file loads alongside the model. llama.cpp's -hf downloads and loads it automatically unless you add --no-mmproj. Calculator with image input on.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 137 GB No | 138 GB No | 142 GB No | 160 GB No | — |
| 3-bit (Q3_K_M) • | 179 GB No | 180 GB No | 185 GB No | 203 GB No | — |
| 4-bit (Q4_K_M) • | 228 GB No | 228 GB No | 233 GB No | 251 GB No | — |
| 5-bit (Q5_K_M) • | 266 GB No | 267 GB No | 272 GB No | 290 GB No | — |
| 6-bit (Q6_K) • | 308 GB No | 308 GB No | 313 GB No | 331 GB No | — |
| 8-bit (Q8_0) • | 398 GB No | 399 GB No | 403 GB No | 421 GB No | — |
| 16-bit (F16/BF16) • | 748 GB No | 749 GB No | 753 GB No | 771 GB No | — |
• real GGUF file size from unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF. Speed is a bandwidth ceiling (1008 GB/s), not a benchmark. Methodology.
Hardware that does fit it
Mac M2 Ultra 192GBMac M3 Ultra 256GBMac M3 Ultra 512GB
What fits instead
- No listed GPU with a list price runs Llama 4 Maverick 17B 128E Instruct fully in memory. The smallest listed hardware that does is the Mac M2 Ultra 192GB (192 GB).
- Largest popular model that fits the RTX 3090 Ti: GLM 4.7 Flash (31B, MoE) at Q4_K_M, about 18.3 GB at 8K context.
"Runs" and "fits" mean the whole model in GPU memory at 8K context (4-bit or better for the model pick). Prices are US launch MSRPs, not today's street prices. "Popular" means the 60 most-downloaded models we track.
More
Llama 4 Maverick 17B 128E Instruct on every GPU and Mac · Every model that runs on the RTX 3090 Ti · Other models that fit 24 GB VRAM · Try other settings in the calculator
Guides: Best GPU for local LLMs by budget · Can I run gpt-oss, Llama 70B or Qwen3? · Mac vs PC for local LLMs
Data: Hugging Face Hub (meta-llama/Llama-4-Maverick-17B-128E-Instruct), NVIDIA specs. Model data last checked 2026-11-18; refreshed hourly.