Phi-4-mini-instruct

Microsoft · 3.8B · ctx 128K · MIT · HuggingFace

Phi-4-mini-instruct (Microsoft) is a 3.8B parameters open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 3.85 GB VRAM for Q4 (weights 2.09 GB + KV 0.76 GB + 1 GB overhead) and 9.36 GB for FP16. About 50 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 1339.7 tok/s (estimated); MI300X 192GB ≈ 887.6 tok/s (estimated); H200 141GB SXM5 ≈ 803.8 tok/s (estimated); H100 80GB SXM5 ≈ 561.0 tok/s (estimated); Gaudi 3 128GB ≈ 548.6 tok/s (estimated). Use the table below for VRAM fit and tok/s per dollar (list MSRP). Estimates follow memory-bandwidth math; measured rows override when present. See methodology for details. Methodology.

100 hardware configs — sorted by tok/s. Default fit context 8K.