- Learn
- Pick a model
gpt-oss 20B vs Qwen3 30B-A3B: two fast MoE models compared
OpenAI's gpt-oss 20B and Qwen3 30B-A3B both run fast on home hardware. Which fits your card, which is smarter where, and estimated speeds.
If you have 16 GB, the choice is made for you: run gpt-oss 20B. With 24 GB or more, try both. Qwen3 30B-A3B usually knows more and writes better; gpt-oss 20B is a sharp, tidy reasoner with lots of room left for context.
Both are mixture-of-experts models. Each holds far more parameters than it uses per token, which is why both feel fast on cards that would crawl through a dense model of the same size.
Side by side
| gpt-oss 20B | Qwen3 30B-A3B | |
|---|---|---|
| Made by | OpenAI | Alibaba (Qwen) |
| Parameters | 21B total, 3.6B active | 30B total, 3B active |
| Download (Ollama) | ~14 GB | ~19 GB |
| Memory needed | ~16 GB | ~21 GB |
| Context | 128K | 32K, 128K with YaRN |
| Reasoning control | Effort: low, medium, high | Thinking on or off |
| Images | No | No |
| License | Apache 2.0 | Apache 2.0 |
| Ollama tag | gpt-oss:20b | qwen3:30b |
Memory is the real difference
Speed is close because each reads about 2 GB of weights per token at the default quant. What differs is how much has to sit in memory: about 16 GB for gpt-oss 20B and about 21 GB for Qwen3 30B-A3B, before you add long context.
So on popular cards:
| Card | gpt-oss 20B | Qwen3 30B-A3B |
|---|---|---|
| RTX 4060 Ti 16 GB | ~50 tok/s | does not fit |
| RTX A4500 20 GB | ~112 tok/s | does not fit |
| RTX 3090 24 GB | ~164 tok/s | ~164 tok/s |
| RTX 5090 32 GB | ~314 tok/s | ~314 tok/s |
| M4 Pro, 64 GB | ~48 tok/s | ~48 tok/s |
All estimates from our calculator. A 20 GB card is the awkward one: gpt-oss fits with room to spare and Qwen3 30B-A3B misses by a gigabyte. On our rack we run the A4500s in pairs for exactly this reason. Two of them measured 110 tok/s on Qwen3.6-35B-A3B, a newer MoE of the same shape.
Where gpt-oss 20B is better
- Structured reasoning. It thinks before every answer, and you choose how hard. Low effort is quick; high effort is slower but catches more on logic and math.
- Tool calling. It was built with agents in mind and is reliable at calling functions in the expected format. See best model for agents and tool calling.
- Long context on 24 GB. It leaves about 8 GB free on a 24 GB card for conversation history and documents.
Where Qwen3 30B-A3B is better
- Knowledge and breadth. More total parameters means it simply knows more, and it shows on general questions and less common topics.
- Writing. gpt-oss can sound flat. Qwen3 is more natural for drafting and rewriting.
- Quick answers. Turn thinking off with
/no_thinkand it replies straight away. gpt-oss always reasons a little first, even at low effort. - Languages. Qwen3 handles many languages well, including Chinese.
Speed in practice
Raw tok/s is nearly the same, but gpt-oss spends tokens reasoning before every reply and Qwen3 only does when you ask. For quick chat, Qwen3 with thinking off often feels faster. For hard problems both spend time thinking, and the fair comparison is how often the answer is right.
Our take
- 16 to 20 GB card: gpt-oss 20B. It's the best model that fits. See best local LLM for 16 GB.
- 24 GB card: Qwen3 30B-A3B as the daily model, gpt-oss 20B when you need long context or an agent. Both download in minutes, so keep both.
- Coding: try both on your own repo. Our coding guide covers the setup.
ollama run gpt-oss:20bRun your own three prompts on each, the way our new model checklist describes. Then check your GPU's page for what else it runs.
One email when there's a new guide, podcast episode, or result from the rack. No spam.