Instructions to use openai/gpt-oss-20b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use openai/gpt-oss-20b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="openai/gpt-oss-20b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")
model = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
AMD Developer Cloud
Local Apps Settings

vLLM

How to use openai/gpt-oss-20b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "openai/gpt-oss-20b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openai/gpt-oss-20b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/openai/gpt-oss-20b

SGLang

How to use openai/gpt-oss-20b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "openai/gpt-oss-20b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openai/gpt-oss-20b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "openai/gpt-oss-20b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openai/gpt-oss-20b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use openai/gpt-oss-20b with Docker Model Runner:
```
docker model run hf.co/openai/gpt-oss-20b
```

Agentic coding test: gpt-oss:20b via Ollama (console 40/40, GUI weak)

#279

by MR1KRAKEN - opened about 12 hours ago

Discussion

MR1KRAKEN

about 12 hours ago

Agentic coding test: gpt-oss:20b (Ollama) — console champion, GUI weak spot

Part of a comparison of five local models run as autonomous coding agents. Overview with the full
matrix: https://huggingface.co/openai/gpt-oss-20b/discussions/278

Russian version of the full report: https://github.com/MR1KRAKEN/local-llm-agent-battle-tests (ru/ folder).
Tested through the Ollama library tag gpt-oss:20b (MXFP4), not through a GGUF from this repo.

Testing was carried out under the supervision of DeepSeek-V4.1-Flash (DSH orchestrator).
It started the server, assigned the task, ran the checks and collected the metrics; it did not write or
fix model code.

Console languages — 40 / 40, the best in the set

Python	C	JavaScript	Rust	Java
8/8	8/8	8/8	8/8	8/8*

All eight reference expressions correct on four languages in a row, 0.1+0.2 → 0.3 included.
* — Java counted only after a separate orchestrator instruction to run javac; the model wrote
Calc.java but never compiled it. Without that intervention the total would be 32/40.

WinForms — no exe in the main run

Metric	Value
Wall time	190 s
Steps	31
Builds	0 / 3 real (+3 calls rejected by the harness)
prompt / completion tokens	21,958 / 7,164
TTFT / speed	1.70 s / 57.12 tok/s
VRAM peak	11,703 MiB
exe	none

Three of 31 steps were burned on invented tool parameters (sandbox_permissions: "", timeoutMs: 0,
empty justification); the harness answered Error: invalid arguments. Files went into an extra
CalculatorBattle\ subfolder, so the first build died with MSB1003. Program.cs was only 8.5 KB —
the implementation was cut off, and the final message was empty.

Compiler errors: CS1001, CS1002, CS1003, CS1022, CS1026, CS1519, CS1525, CS8124 on the first
real build, then CS0117, CS0123, CS0200, CS0246, CS1739 ("no such named argument").

Refinement — from "never built" to working integer arithmetic

It took five rounds, including two separate "build and publish only" tasks, because the model repeatedly
reported success while dist stayed stale.

Round	What the model did	Independent check
1	wrote buttons, styles, `ProcessCmdKey`	exe built, but `dist` not refreshed — edits never reached the binary
2 (build & publish only)	published	exe appeared but was older than `Program.cs`
3	replaced `DataTable.Compute(expr, "$")` with its own parser	mouse `12345=` → `5535`*
4	fixed the grid (`Controls.Add(b, c, r)`), removed `Region` in `OnPaint`	layout correct, clicks land on intended buttons
5 (independent web research, no hints)	fixed parsing and `C`	decimals still broken, `C` does not reset, keyboard operators dead

Root cause found during refinement: the evaluator called dt.Compute(expr, "$"). The second argument
of DataTable.Compute(expression, filter) is a row filter, and "$" is not a valid filter expression —
hence Error on every expression. A small hand-written parser fixed integer arithmetic.

State after refinement (measured externally):

Check	Result
`123*45=` with mouse	`5535` ✅
`46+54=` with mouse	`100` ✅
`2.5+2.5=` with mouse	`27,5` ❌ (second `.` dropped, comma decimal separator)
`C` button	does not reset expression or result ❌
keyboard `2+2⏎`	display `22`, result `Error` — `+` never reaches the handler ❌
window title / grid	`Calculator` ✅ / 4×5 correct ✅

Behaviour worth noting. In the round where it was explicitly asked to research the web and given no
hints, the model never called web_search or web_fetch. Instead it called a non-existent tool named
search seven times (Error: unknown tool "search"), trying to grep its own source for Compute(,
try {, OemComma — seven steps of budget wasted.

MR1KRAKEN

about 11 hours ago

Follow-up: known template issues that affect agent runs

Two template-level reports are worth linking to our numbers:

Discussion #274: the default Jinja template
keeps the commentary channel valid even when no tools are present, and the author measures BFCL_v4 at
~3% on live tests versus 36–60% with a fixed template.
Discussion #272: the template raises on an
assistant message carrying content: null together with tool_calls — exactly what a normal agent loop
produces.

For our run this matters as context rather than as a correction: the console results were measured with tools
present in the request, where the channel bug does not bite. The WinForms failure has a different cause
(invented tool parameters and a never-finished implementation), but template behaviour should be pinned before
comparing scores across harnesses.

Russian version of the full report: https://github.com/MR1KRAKEN/local-llm-agent-battle-tests (folder ru/).

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment