Instructions to use inference-optimization/gpt-oss-1.5B-A1.4B-small with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use inference-optimization/gpt-oss-1.5B-A1.4B-small with Transformers:

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

pipe = pipeline("text-generation", model="inference-optimization/gpt-oss-1.5B-A1.4B-small")
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("inference-optimization/gpt-oss-1.5B-A1.4B-small")
model = AutoModelForCausalLM.from_pretrained("inference-optimization/gpt-oss-1.5B-A1.4B-small", 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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use inference-optimization/gpt-oss-1.5B-A1.4B-small with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "inference-optimization/gpt-oss-1.5B-A1.4B-small"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "inference-optimization/gpt-oss-1.5B-A1.4B-small",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/inference-optimization/gpt-oss-1.5B-A1.4B-small

SGLang

How to use inference-optimization/gpt-oss-1.5B-A1.4B-small 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 "inference-optimization/gpt-oss-1.5B-A1.4B-small" \
    --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": "inference-optimization/gpt-oss-1.5B-A1.4B-small",
		"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 "inference-optimization/gpt-oss-1.5B-A1.4B-small" \
        --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": "inference-optimization/gpt-oss-1.5B-A1.4B-small",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use inference-optimization/gpt-oss-1.5B-A1.4B-small with Docker Model Runner:
```
docker model run hf.co/inference-optimization/gpt-oss-1.5B-A1.4B-small
```

gpt-oss-1.5B-A1.4B-small

This is a small version of unsloth/gpt-oss-20b-BF16 created for testing and development. It keeps all hidden dimensions and hyperparameters of the base model identical, reducing only the number of layers, the layer-type layout, and the number of MoE experts, so it exercises the same architecture (gpt-oss MoE with alternating sliding/full attention, attention sinks, YaRN rope, tanh-limited SwiGLU experts with grouped per-expert tensor storage) at a fraction of the size.

Model Details

Base Model: unsloth/gpt-oss-20b-BF16
Architecture: gpt_oss (GptOssForCausalLM)
Total Parameters: 1.51B (1,510,118,220)
Activated Parameters: 1.41B (1,410,550,860; top-4-of-6 experts per MoE layer)

Configuration Changes

Only the three allowed categories of changes were made (layer counts, layer-type layout, expert count); every other hyperparameter is identical to the base:

Parameter	Base	Small
`num_hidden_layers`	24	2
`layer_types`	24 layers alternating `sliding_attention` / `full_attention`	`sliding_attention`, `full_attention`
`num_local_experts`	32	6

The 2 layers are the minimal set covering every distinct attention layer type (sliding_attention with sliding_window 128, full_attention), mirroring the base's alternating pattern; all base layers are MoE. 6 experts >= num_experts_per_tok 4 (topk+2 for non-degenerate routing). Everything else — hidden size 2880, intermediate size 2880, vocab 201088, 64 Q heads / 8 KV heads / head_dim 64, attention bias + attention sinks, YaRN rope parameters (theta 150000, factor 32, beta_fast 32, beta_slow 1), sliding_window 128, swiglu_limit 7.0, router settings, tie_word_embeddings: false (this base checkpoint has an untied lm_head.weight), max_position_embeddings 131072 — is unchanged from the base.

Checkpoint Structure

Single-file model.safetensors (bf16, ~3.0GB), unsharded. Tensor naming follows the base checkpoint exactly (validated: all 20 tensor-name patterns and both per-layer key sets match the base model.safetensors.index.json modulo layer count), including the gpt-oss "transformer" layout specifics: grouped 3D expert tensors mlp.experts.gate_up_proj / gate_up_proj_bias / down_proj / down_proj_bias, mlp.router.{weight,bias}, self_attn.sinks, and biased q/k/v/o_proj.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("inference-optimization/gpt-oss-1.5B-A1.4B-small", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("inference-optimization/gpt-oss-1.5B-A1.4B-small")

input_ids = tokenizer("According to all known laws", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(output[0]))

Creation Process

This model was created using the llm-compressor create-tiny-model skill.

The config was derived from the base AutoConfig applying only the changes in the table above, the model was built with random init (skip_weights_download + init_weights()), and then fine-tuned on the skill's toy copypasta dataset (8 texts, ~160 steps, lr 1e-4, bf16) until the perplexity stopping criterion was reached.

Validation

Perplexity on the toy dataset text: 1.0001141 (target <= 10)
Greedy generation reproduces the training text ("According to all known laws of aviation, there is no way a bee should be able to fly.")
Checkpoint tensor-name structure matches the base index.json (20/20 patterns; per-layer key sets identical)

Notes

The base unsloth/gpt-oss-20b-BF16 is unquantized bf16 (no quantization_config), so no quantization handling was needed; this small model is plain bf16.
The base has untied embeddings (tie_word_embeddings: false with a separate lm_head.weight in the checkpoint; note the original openai/gpt-oss-20b release is tied, but this unsloth BF16 conversion is not). The untied 201088 x 2880 embedding + lm_head matrices alone cost ~1.16B parameters, which pushes the minimal covering model to 1.51B total.
num_nextn_predict_layers is not present in this architecture; no MTP weights exist.