LocalForge
Back to guide
MODEL / OPENAI

gpt-oss-20b

Use a runtime that applies OpenAI's Harmony response format. Active parameters are not the memory-fit number.

Primary source
Total parameters21B3.6B active / token
Estimated memory~16 GBMXFP4
Context128KText
LicenseApache-2.0Primarily English
01 / BEST FOR
  • reasoning
  • agents
  • tool use
02 / ESTIMATED FIT
Apple silicon · 32 GBComfortableNVIDIA · 24 GBComfortableApple silicon · 64 GBComfortableWorkstation · 48 GBComfortableDatacenter · 80 GBComfortable
03 / DEPLOY WITH
OllamaEasyollama run gpt-oss:20b
llama.cppMediumllama-server -m /path/to/verified-model.gguf --port 8080
LM StudioEasySearch “openai/gpt-oss-20b” → Load a Q4/INT4 build → Start Server
vLLMAdvancedvllm serve openai/gpt-oss-20b --dtype auto
Watch out

Use a runtime that applies OpenAI's Harmony response format. Active parameters are not the memory-fit number.