Three open-weight families now fit on one machine: OpenAI's gpt-oss, Google's Gemma 4 and Alibaba's Qwen3.6. The number that decides whether they run is total parameters, not the active count on the box.
The rule is arithmetic. At 4-bit, a model needs about half a byte per parameter, so 120 billion parameters is roughly 60 GB before the key-value cache. At 8-bit, double it.
| Model | Parameters | Active | About, at 4-bit | Licence | Context |
|---|---|---|---|---|---|
| gpt-oss-20b | 20B | 20B | 10 GB | Apache 2.0 | 128K |
| gpt-oss-120b | 120B | 120B | 60 GB | Apache 2.0 | 128K |
| Gemma 4 31B | 31B | 31B | 16 GB | Apache 2.0 | 262K |
| Qwen3.6-27B | 27B | 27B | 14 GB | Check the card | 262K |
| Qwen3.6-35B-A3B | 35B | 3B | 18 GB | Check the card | 262K |
Dense and mixture-of-experts are different trades
A dense model spends every parameter on every token. A mixture-of-experts model holds all of them in memory and uses a few per token. Qwen3.6-35B-A3B keeps 35 billion parameters resident and runs 3 billion of them, so it needs the memory of a 35B model and answers at the speed of a 3B one.
That is the trade in one line: memory for speed.
What fits where
On a 64 GB machine, gpt-oss-120b is the largest of these that runs comfortably, and 20 GB of it will be left for the cache and the system. Gemma 4 31B and Qwen3.6-27B run with room to spare, and both leave space for a long conversation.
On 32 GB, the 27B and 31B models fit at 4-bit; the 120B does not.
The licence is the other number
gpt-oss ships under Apache 2.0, which permits commercial use without a separate agreement. Gemma 4 is also Apache 2.0. Qwen's cards carry their own terms, and they change between releases, so read the card for the version you download rather than the family name.
Speed on your own hardware is a measurement, not a claim. Load the model, time 500 output tokens, and use that number.