Llama 3.1 8B
8BOllamaLlama · Confidence: high
Model memory5.7 / 22.1 GB
- Estimated speed
- 92–139.2 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 12 GB
ollama run llama3.1:8b-q4_K_MLOCAL AI COMPATIBILITY
AI Rig Match
Match your GPU or Apple silicon Mac with local language, coding, image, and video models.
Capacity and speed are modeled estimates, not a guarantee. Confirm runtime, driver, and model-version requirements before buying hardware.
YOUR RIG
Build a hardware profile
Set the memory budget and workload that matter for this run.
MATCH REPORT
Compatible models
Results separate native VRAM fit from slower system-memory offload.
5Native
2Offload
0Upgrade
Qwen3 8B
8BOllamaQwen · Confidence: high
Model memory6.1 / 22.1 GB
- Estimated speed
- 87.4–132.2 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 12 GB
ollama run qwen3:8b-q4_K_MGemma 3 12B
12BOllamaGemma · Confidence: high
Model memory8.9 / 22.1 GB
- Estimated speed
- 59.3–89.7 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 16 GB
ollama run gemma3:12b-q4_K_MDeepSeek R1 Distill 14B
14BOllamaDeepSeek · Confidence: high
Model memory10.1 / 22.1 GB
- Estimated speed
- 49.9–75.5 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 20 GB
ollama run deepseek-r1:14b-q4_K_MGPT-OSS 20B
20BOllamaOpenAI · Confidence: medium
Model memory14 / 22.1 GB
- Estimated speed
- 38.2–57.8 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 24 GB
ollama run gpt-oss:20b-q4_K_MQwen3 32B
32BOllamaQwen · Confidence: medium
Model memory22.3 / 22.1 GB
- Estimated speed
- 7.3–11 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 40 GB
ollama run qwen3:32b-q4_K_MLlama 3.3 70B
70BOllamaLlama · Confidence: medium
Model memory46.2 / 22.1 GB
- Estimated speed
- 3.7–5.7 tok/s
- Quantization
- Q4
- Context
- 8K
- Minimum RAM
- 64 GB
ollama run llama3.3:70b-q4_K_MRanges include runtime overhead and an OS reserve. Real performance varies by model build, backend, driver, thermals, and prompt.
CALCULATION METHOD
A fit result with its assumptions exposed
The estimate treats weights, context, runtime overhead, and memory topology as separate constraints.
Weights before marketing labels
Each profile starts from a quantized weight footprint and adds runtime headroom instead of equating parameter count with VRAM.
Context has a real cost
Longer context adds KV-cache pressure. The same model can move from native fit to offload when context grows.
Offload is not native fit
System-memory offload is shown separately because it can run while delivering a very different speed and latency experience.
COMMON QUESTIONS
Local AI hardware questions
How much VRAM does an 8B language model need?
A Q4 build commonly needs about 5–7 GB after basic runtime overhead. Longer context, a larger batch, or higher precision can push it beyond 8 GB.
Is Apple unified memory the same as GPU VRAM?
No. Apple silicon shares one memory pool between the CPU and GPU. That flexibility helps larger models fit, but macOS and other processes still need a reserve.
What does offload mean?
Part of the model stays in system RAM and moves through the CPU or PCIe path. It can make a model launch, but it is usually much slower than keeping all active data in GPU memory.
Are the speed ranges benchmark results?
They are conservative modeled ranges, not a benchmark claim. Exact speed depends on runtime, kernel support, model build, driver, cooling, and workload.