# Can I run Llama 3.3 70B on dual RTX 3090s? | What Can I Run? > No. Llama 3.3 70B needs about 50.2 GB at Q4_K_M, and dual RTX 3090s has 48.0 GB usable. What to run instead. Source: https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/dual-rtx-3090/ [Home](https://whatcanirunai.vercel.app/) / [Can I run it?](https://whatcanirunai.vercel.app/can-i-run/) / [Llama 3.3 70B](https://whatcanirunai.vercel.app/models/llama-3.3-70b-instruct/) / Llama 3.3 70B on dual RTX 3090s **No.** Llama 3.3 70B at Q4\_K\_M needs about 50.2 GB, and dual RTX 3090s have 48.0 GB usable at a \~4k context. # Can I run Llama 3.3 70B on dual RTX 3090s? Llama 3.3 70B at Q4\_K\_M needs about 50.2 GB, and dual RTX 3090s have 48.0 GB usable. Two cards add memory, not decode speed; the speed estimate uses one RTX 3090’s 936 GB/s. Meta’s flagship 70B instruct model, the usual target for a high-end local chat box. These are estimates from [our formula](https://whatcanirunai.vercel.app/method/), not measured benchmarks. Quant that fits None of Q4, Q8 or FP16 needs 50.2 GB Memory needed 50.2 GB at Q4\_K\_M, \~4k context Usable VRAM 48.0 GB of 48 GB VRAM Est. tokens/sec n/a does not fit [Llama-3.3-70B-Instruct](https://whatcanirunai.vercel.app/models/llama-3.3-70b-instruct/) · [GeForce RTX 3090](https://whatcanirunai.vercel.app/hardware/rtx-3090/) · [Multi-GPU](https://whatcanirunai.vercel.app/multi-gpu-for-local-llm/) · [Dual 3090 build](https://whatcanirunai.vercel.app/builds/dual-used-rtx-3090-70b/) · [Everything this machine runs](https://whatcanirunai.vercel.app/what-can-i-run/rtx-3090/) · [Q4 vs Q8 vs FP16](https://whatcanirunai.vercel.app/quantization/) ## Memory at each quant | Quant | Need vs 48.0 GB | Verdict | | -------- | --------------- | ------------- | | Q4\_K\_M | 50.2 GB | Q4Won’t fit | | Q8\_0 | 87.2 GB | Q8Won’t fit | | FP16 | 163.3 GB | FP16Won’t fit | ## Estimated tokens per second No speed estimate: Llama 3.3 70B does not fit in the 48.0 GB this machine can use, so we do not publish a tokens/sec figure for CPU offload. ## Context headroom Even at a short \~4k context, the weights do not fit. Extending context would only increase the gap. __Context headroom for Llama 3.3 70B on dual RTX 3090s at Q4\_K\_M__ | Context | KV cache (FP16) | Total at Q4\_K\_M | Verdict | | ------- | --------------- | ----------------- | --------- | | 4k | 1.3 GB | 50.2 GB | Won’t fit | | 8k | 2.7 GB | 51.5 GB | Won’t fit | | 16k | 5.4 GB | 54.2 GB | Won’t fit | | 32k | 10.7 GB | 59.6 GB | Won’t fit | | 64k | 21.5 GB | 70.3 GB | Won’t fit | | 128k | 42.9 GB | 91.8 GB | Won’t fit | From the model’s config. Try other sizes in the [VRAM calculator](https://whatcanirunai.vercel.app/vram-calculator/) or see the [context length guide](https://whatcanirunai.vercel.app/context-length/). ## What to do instead * Run [R1 Distill 32B](https://whatcanirunai.vercel.app/models/deepseek-r1-distill-qwen-32b/) instead (23.8 GB at Q4). It is a different model, not a drop-in, but it does fit this machine. * To hold Llama 3.3 70B at Q4, you need about 50.2 GB. The [MacBook Pro M5 Max 128GB](https://whatcanirunai.vercel.app/hardware/macbook-pro-m5-max-128/) has 96.0 GB usable ($6699 Apple list price, 14-inch M5 Max 18C CPU/40C GPU, 128GB/2TB). * Our estimate is only 2.2 GB over this machine’s usable memory, so a tighter quant or a shorter context may work in practice. We still mark it as a no under [the published formula](https://whatcanirunai.vercel.app/method/). * CPU offload can load the weights anyway, but it is slow. We do not publish an offload tokens/sec figure. * See the [best GPU for this model family](https://whatcanirunai.vercel.app/best-gpu-for/llama/) and the [Llama-3.3-70B-Instruct model page](https://whatcanirunai.vercel.app/models/llama-3.3-70b-instruct/). ## Other hardware for Llama 3.3 70B * [Llama 3.3 70B on an RTX 3060 12GB](https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/rtx-3060-12gb/)Won’t fit * [Llama 3.3 70B on an RTX 3070 Ti](https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/rtx-3070-ti/)Won’t fit * [Llama 3.3 70B on an RTX 3080](https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/rtx-3080-10gb/)Won’t fit * [Llama 3.3 70B on an RTX 3090](https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/rtx-3090/)Won’t fit * [Llama 3.3 70B on an RTX 4060 Ti 16GB](https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/rtx-4060-ti-16gb/)Won’t fit * [Llama 3.3 70B on an RTX 4070 Ti Super](https://whatcanirunai.vercel.app/can-i-run/llama-3.3-70b-instruct/rtx-4070-ti-super/)Won’t fit ## Other models on dual RTX 3090s * [R1 Distill 70B on dual RTX 3090s](https://whatcanirunai.vercel.app/can-i-run/deepseek-r1-distill-llama-70b/dual-rtx-3090/)Won’t fit * [gpt-oss-120b on dual RTX 3090s](https://whatcanirunai.vercel.app/can-i-run/gpt-oss-120b/dual-rtx-3090/)Won’t fit * [Qwen3-235B on dual RTX 3090s](https://whatcanirunai.vercel.app/can-i-run/qwen3-235b-a22b/dual-rtx-3090/)Won’t fit * [DeepSeek-R1 on dual RTX 3090s](https://whatcanirunai.vercel.app/can-i-run/deepseek-r1/dual-rtx-3090/)Won’t fit * [Mixtral 8x22B on dual RTX 3090s](https://whatcanirunai.vercel.app/can-i-run/mixtral-8x22b-instruct-v0.1/dual-rtx-3090/)Won’t fit * [GLM-4.5-Air on dual RTX 3090s](https://whatcanirunai.vercel.app/can-i-run/glm-4.5-air/dual-rtx-3090/)Won’t fit ## Quick answers Can I run Llama 3.3 70B on dual RTX 3090s? No. Llama 3.3 70B at Q4\_K\_M needs about 50.2 GB, and dual RTX 3090s have 48.0 GB usable. How much VRAM does Llama 3.3 70B need? About 50.2 GB at Q4\_K\_M, 87.2 GB at Q8\_0 and 163.3 GB at FP16, estimated for a \~4k context. How fast is Llama 3.3 70B on dual RTX 3090s? We do not estimate tokens/sec when the model does not fit. CPU offload is possible but slow. What context length can I use? The weights already do not fit, so longer context is not the limiting factor. --- Agent guide: https://whatcanirunai.vercel.app/llms.txt · JSON API: https://whatcanirunai.vercel.app/developers/