The direct answer
Yes, you can run Llama 3 on 8GB — but only the 8B parameter version at Q4 quantization. Llama 3.1 8B at Q4_K_M uses approximately 5.5 GB of VRAM with 4K context, leaving 2.5 GB for your system. This runs on any 8 GB GPU (RTX 3070, RTX 4060, RTX 2080) and on Apple Silicon Macs with 8 GB unified memory.
No, you cannot run Llama 3.3 70B on 8GB. The 70B model at minimum quantization (Q2) needs ~18 GB. At Q4 it needs ~35 GB. Even with aggressive CPU offloading, generation speed drops to <1 tok/s, which is unusable.
Here is the full breakdown of what fits and what does not.
What fits on 8GB?
These models run comfortably on 8 GB VRAM or unified memory at the specified quantization levels:
| Model | Quantization | VRAM usage | Context | tok/s (GPU) |
|---|---|---|---|---|
| Llama 3.1 8B | Q4_K_M | ~5.5 GB | 4K | ~20-25 |
| Llama 3.2 3B | Q4_K_M | ~2.5 GB | 8K | ~30-40 |
| Llama 3.2 1B | Q4_K_M | ~1 GB | 8K | ~40-60 |
| Qwen 2.5 7B | Q4_K_M | ~5 GB | 8K | ~20-28 |
| Mistral 7B | Q4_K_M | ~4.5 GB | 8K | ~25-30 |
| Phi-3 3.8B | Q4_K_M | ~2.8 GB | 4K | ~30-40 |
Notice that Llama 3.1 8B at Q4 is the largest Llama model you can fit on 8 GB. The smaller Llama 3.2 models (1B and 3B) are faster but less capable.
What does NOT fit on 8GB?
| Model | Minimum VRAM | Why it does not fit |
|---|---|---|
| Llama 3.3 70B Q4 | ~35 GB | Weights alone are 35 GB + KV cache. Needs 48 GB total. |
| Llama 3.3 70B Q2 | ~18 GB | Even the most aggressive quantization does not fit on 8 GB. |
| Qwen 2.5 32B Q4 | ~18 GB | ~18 GB weights + context. Needs 24 GB GPU. |
| Mixtral 8x7B Q4 | ~20 GB | ~20 GB weights + context. Needs 24 GB GPU. |
| Llama 3.1 8B Q8 | ~8.5 GB | Only if context is minimal (1K). At 4K context it hits ~10 GB. |
The Llama 3.3 70B family requires 48 GB+ of VRAM — out of reach for any single consumer GPU under $2000. You need dual RTX 3090s (48 GB total) or a Mac with 64 GB+ unified memory.
Performance on 8GB
Your experience depends on whether you are using a GPU, Apple Silicon, or CPU-only:
| Hardware | Model | tok/s | Usability |
|---|---|---|---|
| RTX 3070 8 GB | Llama 3.1 8B Q4 | ~20-25 | Great — fast enough for interactive chat |
| RTX 4060 8 GB | Llama 3.1 8B Q4 | ~18-22 | Great — slightly slower than 3070 |
| M1 MacBook Air 8 GB | Llama 3.1 8B Q4 | ~10-15 | Good — usable, fans stay off for short sessions |
| CPU-only 8 GB RAM | Llama 3.1 8B Q4 | ~2-5 | Slow — usable for batch processing, not chat |
| CPU-only 8 GB RAM | Llama 3.2 3B Q4 | ~5-8 | OK — slow but usable for simple tasks |
The threshold for "usable" interactive chat is ~10 tok/s. Below that, the delay between pressing Enter and seeing the response feels sluggish. Above 20 tok/s, text appears smoothly as if you are reading a fast typist.
How to set up on 8GB hardware
Follow these steps to run Llama 3.1 8B on your 8 GB machine:
Step 1: Check your VRAM
First, confirm you have enough free VRAM. Close other GPU-using applications (browser with hardware acceleration, design tools, games).
# NVIDIA GPUs nvidia-smi --query-gpu=memory.free --format=csv,noheader # Apple Silicon (unified memory) memory_pressure # General (system RAM) free -h
Step 2: Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
Step 3: Run Llama 3.1 8B
# Default (Q4, 8K context) ollama run llama3.1:8b # If you hit memory limits, reduce context: ollama run llama3.1:8b --context-size 4096 # For the tiny Llama 3.2 3B (faster, less capable): ollama run llama3.2:3b
Ollama automatically selects the best quantized version for your hardware. For llama.cpp, use:
./main -m llama-3.1-8b-q4_k_m.gguf -n 512 -ngl 99 -c 4096
The -ngl 99 flag offloads all layers to GPU. If you run out of memory, lower -ngl to offload fewer layers (e.g., -ngl 20 offloads ~80% to GPU, rest to CPU).
Quantized vs full quality comparison
How much quality do you lose by running Llama 3 at Q4 instead of FP16?
| Quantization | Size vs FP16 | Quality retention | Use case |
|---|---|---|---|
| FP16 (full) | 100% | 100% (baseline) | Reference, benchmarking |
| Q8 | 50% | ~99.5% | Maximum quality, plenty of VRAM |
| Q6 | 37.5% | ~98.5% | Good balance for larger models |
| Q4 | 25% | ~95-97% | Standard for consumer GPUs |
| Q3 | 18.75% | ~90-93% | Fitting very large models |
| Q2 | 12.5% | ~80-85% | Compressed models, quality loss visible |
In blind A/B tests, most users cannot consistently distinguish Q4 from FP16 on chat, summarisation, and Q&A tasks. The difference is noticeable on sensitive tasks like translation of idioms, creative writing nuance, and exact mathematical reasoning. For 95% of everyday use, Q4 is indistinguishable from full precision.
Experiment with our quantization calculator to compare quality vs VRAM trade-offs for any model.
How to check your hardware
Before downloading models, verify your hardware specs:
- GPU VRAM: Run
nvidia-smi(NVIDIA) or check About This Mac → System Report → Graphics/Displays (Apple Silicon). - System RAM: Run
free -h(Linux), check Activity Monitor (Mac), or Task Manager (Windows). - GPU compute capability: Ensure your GPU supports CUDA (NVIDIA cards from GTX 900 series and up work).
- Available disk space: GGUF model files are 3-8 GB. Ensure you have at least 20 GB free for a few models.
Use our VRAM calculator to check exactly which model + quantization combinations fit before you download.
Alternatives to running on 8GB hardware
If 8 GB is too tight for your needs, you have options:
- Upgrade your GPU: A used RTX 3060 12 GB costs ~$200 and doubles your VRAM. An RTX 3090 24 GB is ~$700 used.
- Use CPU offloading: llama.cpp can split layers between GPU and system RAM. Expect ~5-10 tok/s with 50% offloading.
- Use a cloud API: If you need Llama 3 70B or GPT-4-class models, skip the hardware entirely. Plugsky gives you 38 models including all Llama versions with no hardware requirements.
Cannot fit Llama 70B on your hardware?
Plugsky gives you 38 models including Llama 3.3 70B, Qwen 2.5 72B, and more. No hardware, no setup, no rate limits.
Start Free → VRAM calculatorFrequently asked questions
Can I run Llama 3 on 8GB RAM?
Yes. Llama 3.1 8B at Q4 quantization uses ~5.5GB of VRAM with 4K context. It runs comfortably on any 8GB GPU (RTX 3070, RTX 4060, etc.) and on Apple Silicon with 8GB unified memory at ~10-15 tok/s.
Can I run Llama 3 70B on 8GB?
No. Llama 3.3 70B at minimum quantization (Q2) still needs ~18GB VRAM. At Q4 it needs ~35GB. The 70B model is completely out of reach for 8GB hardware. You need 48GB+ (dual GPU or Mac with 64GB+).
How fast will Llama 3 run on 8GB?
On an 8GB GPU (e.g., RTX 3070): ~20-25 tok/s for Llama 3.1 8B Q4. On 8GB Apple Silicon (e.g., M1 Air): ~10-15 tok/s. On 8GB system RAM with CPU only: ~2-5 tok/s. Only GPU inference provides usable speeds.
Is quantized Llama 3 as good as the full version?
Q4 quantization retains ~95-97% of the original model quality. In blind tests, most users cannot tell the difference between Q4 and FP16 for chat and summarisation. Q8 retains ~99.5%. The quality loss is minimal and well worth the 4x memory savings.
How do I run Llama 3 on 8GB hardware?
Install Ollama (curl -fsSL https://ollama.com/install.sh | sh), then run 'ollama run llama3.1:8b'. Ollama automatically downloads the optimal Q4 quantized version. For more control, use llama.cpp with the --ngl flag to offload layers to GPU.