Can I run GPT-OSS 20B?

GPT-OSS 20B is OpenAI's open-weight reasoning model, and the key thing to understand is that it is a mixture-of-experts: of its 20.9B total parameters, only 3.6B activate per token. That means it runs much faster than a dense 20B model would, but you still have to hold the whole thing in memory. At a 4-bit quant it lands around 12.7 GB, and the practical floor is about 24 GB of RAM. It does not fit on a 12 GB card like an RTX 3060, so realistically you are looking at a 24 GB GPU like a 4090, or an Apple Silicon Mac with plenty of unified memory.

In daily use it feels quick for its size, which is the MoE payoff. On an RTX 4090 you can expect around 393 tokens per second, and on an M-series Max chip roughly 160 tok/s, both well past reading speed. CPU-only on DDR5 drops to about 23 tok/s, usable for batch jobs but not interactive chat. The 128K context window is generous, but it is the part that quietly eats memory: fill it all the way and total usage climbs to about 38.5 GB once the KV cache is loaded. Keep working context modest unless you have headroom to spare, or drop to the q2 build near 8.8 GB.

It is built for chat and reasoning rather than coding, so for programming work something like Codestral 22B generally serves you better, and Mistral Small 3.1 24B is the pick if you also want vision. Its standout trait is reasoning quality at a 3.6B active-parameter cost, and if you want more headroom the larger GPT-OSS 120B sibling scales the same recipe up. The license is Apache 2.0, so it is genuinely free to use, including commercially and in production, with no provider-specific strings attached.