← CanIRunLocal · All hardware · All models

Can I run Jais 3 7B locally?

IFM · 9B parameters · chat, reasoning.

Ollama weight files: Q4_K_M 5.59 GB (hf.co/IFM/Jais-3-7B-GGUF:Q4_K_M, size from the Hugging Face file listing); Q5_K_M 6.47 GB (hf.co/IFM/Jais-3-7B-GGUF:Q5_K_M, size from the Hugging Face file listing); Q8_0 9.57 GB (hf.co/IFM/Jais-3-7B-GGUF:Q8_0, size from the Hugging Face file listing); BF16 18.01 GB (hf.co/IFM/Jais-3-7B-GGUF:BF16, size from the Hugging Face file listing).

Yes on 82 machines in this list. The smallest is RTX 3060 8GB: Q5_K_M, about 7.5 GB, roughly 22 tokens per second, via Ollama. ollama run hf.co/IFM/Jais-3-7B-GGUF:Q5_K_M

• DGX Spark 128GB: BF16, ~19.1 GB, fits in fast memory, ~11 tok/s — ollama run hf.co/IFM/Jais-3-7B-GGUF:BF16
• Mac Studio M5 Max 128GB: BF16, ~19.1 GB, fits in fast memory, ~27 tok/s — ollama run hf.co/IFM/Jais-3-7B-GGUF:BF16
• Mac Studio M5 Ultra 256GB: BF16, ~19.1 GB, fits in fast memory, ~53 tok/s — ollama run hf.co/IFM/Jais-3-7B-GGUF:BF16

Setup on DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte)

• Ollama: ollama run hf.co/IFM/Jais-3-7B-GGUF:BF16
• LM Studio: lms get IFM/Jais-3-7B-GGUF@Q8_0
• llama.cpp: llama-server -hf IFM/Jais-3-7B-GGUF:Q8_0 -c 8192
• vLLM: vllm serve inception42/Jais-3-7B --max-model-len 8192 --gpu-memory-utilization 0.9

Hardware that fits

Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-20.

A project by OnDevice Labs · llms.txt