GPT-OSS-120B-F16

120B MoE model.

llama.cpp Mixed Cross-platform General Updated 2026-06-05

Import

Requires llml ≥ 0.5.0

$ llml import https://llml.dev/profiles/GPT-OSS-120B-F16.toml --activate

$ brew install --cask flyingnobita/tap/llml

$ llml import https://llml.dev/profiles/GPT-OSS-120B-F16.toml --activate

$ go install github.com/flyingnobita/llml/cmd/llml@latest

$ llml import https://llml.dev/profiles/GPT-OSS-120B-F16.toml --activate

$ scoop bucket add flyingnobita https://github.com/flyingnobita/scoop-bucket && scoop install flyingnobita/llml

$ llml import https://llml.dev/profiles/GPT-OSS-120B-F16.toml --activate

Download .toml

Copy the Run command, paste it in your terminal, and pick a local model. The profile attaches immediately — press p in the TUI to confirm.

Why this profile exists

120B MoE model; F16 quant ~66GB unified memory for 6+ tokens/s with 1-bit quant; -ot regex offloads all MoE expert layers to CPU, fitting non-MoE layers on a single GPU; adjust regex for partial offload; recommended ctx minimum 16384, max 131072

Launch configuration

# args

--ctx-size 16384

--n-gpu-layers 99

-ot ".ffn_.*_exps.=CPU"

--temp 1.0

--min-p 0.0

--top-p 1.0

--top-k 0

Hardware assumptions

Mixed — tested envelope
Cross-platform — backend installed and on PATH
Backend: llama.cpp >= current llml-supported version
Profile assumes the model file is already on disk; llml supplies the path at launch

GPT-OSS-120B-F16

Why this profile exists

Launch configuration

Hardware assumptions

Keyboard shortcuts