gpt-oss-20b
gpt-oss-20b is OpenAI’s small open model from 2025, a mixture of experts with 3.6 billion active parameters. It handles text only, has adjustable reasoning effort and calls tools reliably. Quantised it needs about 13 GB and runs quickly on a computer with 32 GB.
Profile
- Maker
- OpenAI · United States
- Licence
- Free for commercial use
- Runs on
- Computer with 32 GB
- Memory at 4 bits
- 13 GB
- Memory at 8 bits
- 12 GB
- Parameters
- 20.9bn, 3.6bn active
- Context
- 131,000 tokens
- Inputs
- text
- Released
- August 2025
Scores
- Human rating
- 1287 · rank 109 of 226 open
- In German
- 1255 · rank 91 of 154 open
- Epoch index
- 137.8 · rank 44 of 127 open
- GPQA Diamond (expert questions)
- 47.7 %
- AIME (maths)
- 65.2 %
- LMArena · Coding
- 1307 · 105/221
- LMArena · Maths
- 1317 · 90/215
- LMArena · Creative writing
- 1201 · 133/224
- LMArena · Following instructions
- 1236 · 122/226
- LMArena · Hard prompts
- 1274 · 110/226
- LMArena · Long prompts
- 1250 · 121/205
- LMArena · Multi-turn dialogue
- 1268 · 115/224
Scores: LMArena, Epoch AI, EuroEval, Hugging Face, models.dev; as of 19 October 2026. Sources and method →
Use cases
Where the model makes sense.
Data from texts as JSON
- Situation
- Fixed fields from orders, requests or minutes should be transferred into another system.
- What it looks like
- The model gets a schema and returns a JSON object with the fields for each text, which is processed directly.
- What to watch
- Check required fields and formats automatically after output and pass errors to a person.
Prototypes for AI features
- Situation
- A development team wants to try an AI feature before spending money on a cloud API.
- What it looks like
- The prototype runs locally with gpt-oss-20b via an OpenAI-compatible interface; the model can be swapped later.
- What to watch
- Compare results with the later target model; a prototype with a small model often underestimates quality.
Plausibility checks
- Situation
- Quotes, timesheets or travel expenses should be checked for anomalies before approval.
- What it looks like
- With higher reasoning effort the model recalculates, compares with rules and flags unusual items.
- What to watch
- Leave arithmetic to a tool such as a spreadsheet and use the model to explain.
Agents with tools
- Situation
- An internal assistant should look up appointments, stock levels or ticket status.
- What it looks like
- The model decides which tool to call and summarises the answer. Tools are connected via MCP.
- What to watch
- Allow write tools only with confirmation by a person.
The use cases were written with AI assistance (Claude by Anthropic) and reviewed by vona. Scores come unchanged from the sources listed.
When another model fits better
It is not a good fit for customer-facing German texts or images. Gemma 4 31B or Mistral Small 3.2 in the same hardware class are better suited.
How to start
ollama run gpt-oss:20bWith Ollama or LM Studio on a computer with 32 GB. Reasoning effort (low, medium, high) can be chosen per request.
Suitable interfaces and programs from our directory: Ollama, LM Studio, LibreChat
Licence: Free for commercial use
The licence allows commercial use, modification and redistribution; the main obligation is to keep the licence notice.
Hugging Face ↗Similar models
Comparable models.
OpenAI · United States
gpt-oss-120b
OpenAI's open text model from 2025 with adjustable reasoning effort and tool calls. Fits on a Mac Studio with 128 GB or a single 80 GB GPU.
Free for commercial use · Runs on: Mac Studio with 128 GB
Use cases and details →Alibaba (Qwen) · China
Qwen 3.8 27B
Compact model from the Qwen 3.8 series that understands images and calls tools. Quantised, it runs on a computer with 32 GB and reaches scores otherwise seen from much larger models.
Free for commercial use · Runs on: Computer with 32 GB
Use cases and details →NVIDIA · United States
Nemotron 3 Nano 30B-A3B
Small NVIDIA mixture-of-experts text model with three billion active parameters, built for fast agents and tool calls on your own hardware.
Commercial use with conditions · Runs on: Workstation with 64 GB
In the overview →
FAQ
What hardware do I need for gpt-oss-20b?
In the common 4-bit version, gpt-oss-20b needs about 13 GB of memory including headroom for context and runtime. That fits: Computer with 32 GB. The more accurate 8-bit version needs about 12 GB. Long documents increase the requirement.
May I use gpt-oss-20b commercially?
The licence allows commercial use, modification and redistribution; the main obligation is to keep the licence notice.
How good is gpt-oss-20b in German?
On LMArena, people rate gpt-oss-20b at 1255 points for prompts in German, rank 91 of 154 open models. The best way to see whether it suits your texts is a test with your own examples.
AI consulting
Your own language model for your company?
Together we find the model that fits your tasks, your hardware and your data protection requirements, test it with real documents and set up operation.