Running Llama 4 Maverick locally is not a casual project. This is a 400B MoE model — 400 billion total parameters with ~17B active per token. At FP16, the weights alone need ~800GB of memory. Even at aggressive Q4 quantization, you are looking at 200-250GB. No single consumer GPU comes close. So the real question is: which multi-GPU setup or cloud option makes Maverick practical, and is it even worth self-hosting?
NVIDIA GeForce RTX 4090
24GB GDDR6X24GB VRAM per card. Four RTX 4090s give 96GB, which runs Maverick only with heavy CPU offload — Q4 alone is ~245GB.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
VRAM requirements at a glance
| Quantization | Total VRAM needed | Minimum hardware |
|---|---|---|
| FP16 | ~800GB | 8x A100 80GB or 10x H100 |
| Q8_0 | ~400GB | 8x RTX 4090 or 4x A100 80GB |
| Q4_K_M | ~245GB | 4x A100 80GB — 96GB of consumer cards needs CPU offload |
| Q3_K_M | ~170-200GB | 3x RTX 4090 or 2x A6000 (96GB) |
| Q2_K | ~130-150GB | 2x RTX 4090 (48GB) + CPU offload |
These numbers include model weights only. Maverick costs about 0.19MB per token of KV cache at FP16, so add roughly 6GB at 32K context, 24GB at 128K, and 192GB at the full 1M — the cache outgrows the weights long before you reach the advertised context length. Maverick supports 1M token context in theory, but consumer setups are limited to 4K-8K context to keep total memory usage manageable.
Scenario 1: Cheapest self-hosting path
Dual RTX 4090s (~$4,400 total)
Two RTX 4090s give you 48GB of combined VRAM. That is not enough for Q4 (~200GB+), so you would need aggressive Q2-Q3 quantization plus significant CPU offloading. The result:
- Q2_K weights: ~130GB, split between 48GB GPU + 82GB system RAM
- Performance: ~5-8 tok/s with heavy CPU offload
- Quality: noticeably degraded at Q2
Honestly? We do not recommend this approach for daily use. The quality loss at Q2 and the slow inference speed from CPU offloading make it more of a proof-of-concept than a practical setup. You are better off using a smaller model like Llama 4 Scout on a rented 80GB card or a four-card 3090 build, where you get good speed and quality at a fraction of the memory.
Scenario 2: Serious self-hosting
Four RTX 4090s (~$8,800 total)
Four cards give you 96GB of VRAM. Maverick at Q3_K_M (~180GB) still needs CPU offloading, but with 96GB on GPUs and the rest in RAM, performance is reasonable:
- Q3_K_M: ~12-15 tok/s with partial CPU offload
- Q4_K_M with heavy offloading: ~8-12 tok/s
You also need a motherboard with 4 x16 PCIe slots, a 1600W+ PSU, and a case that can cool four 450W GPUs. This is a dedicated workstation build costing $8,000-10,000 all-in.
Two NVIDIA A6000s (~$7,000 total)
Each A6000 has 48GB VRAM (96GB combined). Same total VRAM as four 4090s but in two cards — simpler cooling, fewer PCIe slots, less PSU pressure. The A6000 has lower memory bandwidth than the 4090, so expect ~10-14 tok/s at Q3. Professional cards also hold value better on the resale market.
Check NVIDIA RTX A6000 48GB on Amazon→Buy on Shopee SG→Scenario 3: Cloud hosting (the pragmatic choice)
For most users, cloud is the right answer for Maverick. The economics are straightforward:
| Cloud setup | Cost | Performance |
|---|---|---|
| RunPod 8x A100 80GB | ~$12-24/hr | ~40-60 tok/s at Q4 |
| RunPod 4x A100 80GB | ~$6-12/hr | ~25-35 tok/s at Q4 |
| Vast.ai 4x A100 | ~$4-8/hr | ~25-35 tok/s at Q4 |
At $12/hr for a fast 8x A100 setup, you would need to use Maverick for 500+ hours to justify the $8,800 cost of four RTX 4090s — and the cloud setup is faster, requires zero maintenance, and scales instantly.
Try Vast.ai Cloud GPU→Our recommendation: Use cloud for Maverick unless you have specific privacy requirements or need 24/7 availability without recurring costs. The self-hosting math does not work out for occasional use.
Scenario 4: API access (cheapest of all)
If you do not need to self-host for privacy or customization reasons, API access through providers like Together, Fireworks, or Groq costs a fraction of self-hosting. At $0.50-1.00 per million tokens, you would need to process millions of tokens daily before self-hosting becomes cheaper.
We mention this because many users jump to “I need to run it locally” without considering the cost comparison. For Maverick specifically, the API route saves thousands of dollars per year.
Hardware checklist for self-hosting
If you are committed to running Maverick locally, here is what you need:
| Component | Minimum | Recommended |
|---|---|---|
| GPUs | 2x RTX 4090 (48GB) | 4x RTX 4090 (96GB) or 2x A6000 |
| System RAM | 128GB DDR5 | 256GB DDR5 |
| PSU | 1200W | 1600W (for 4 GPUs) |
| Motherboard | 2x PCIe x16 | 4x PCIe x16 (HEDT/server) |
| Storage | 500GB NVMe | 1TB NVMe (for model files + swap) |
| CPU | Any modern 8-core | Threadripper for PCIe lanes |
System RAM matters. CPU offloading uses system memory for layers that do not fit on GPUs, and DDR5 bandwidth directly affects offloaded layer performance. 128GB is the minimum; 256GB gives headroom for larger context windows.
Check NVIDIA GeForce RTX 4090 on Amazon→Buy on Shopee SG→Should you even run Maverick?
Bluntly: most people should not self-host this model. The 17B active parameters deliver excellent quality, but you can get similar-tier reasoning from smaller models that fit on a single GPU:
- Qwen 3 32B — fits on one RTX 4090, comparable quality for many tasks
- Llama 4 Scout — 109B MoE, ~67GB at Q4, so an 80GB card or four 24GB cards; strong performance
- Gemma 4 31B Dense — single-GPU model with competitive benchmarks
Maverick’s advantage is its massive expert pool and 1M context — if you do not need those, a smaller model is more practical.
Check NVIDIA GeForce RTX 5090 on Amazon→Buy on Shopee SG→For the Llama 4 family, see our best GPU for Llama 4 overview and the Llama 4 Scout GPU guide. VRAM planning across the Llama 4 lineup is covered in how much VRAM for Llama 4. Building a multi-GPU rig? Check our best multi-GPU setup for LLM guide.