Instructions to use allenai/OLMo-2-0425-1B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use allenai/OLMo-2-0425-1B with Transformers:

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

pipe = pipeline("text-generation", model="allenai/OLMo-2-0425-1B")

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

tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-2-0425-1B")
model = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-0425-1B", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use allenai/OLMo-2-0425-1B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "allenai/OLMo-2-0425-1B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "allenai/OLMo-2-0425-1B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/allenai/OLMo-2-0425-1B

SGLang

How to use allenai/OLMo-2-0425-1B 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 "allenai/OLMo-2-0425-1B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "allenai/OLMo-2-0425-1B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "allenai/OLMo-2-0425-1B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "allenai/OLMo-2-0425-1B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use allenai/OLMo-2-0425-1B with Docker Model Runner:
```
docker model run hf.co/allenai/OLMo-2-0425-1B
```

Main revision

by aldakata - opened Oct 7, 2025

Discussion

aldakata

Oct 7, 2025

Hi!
I was playing around with the revisions and I get different results with the main and the stage2-ingredient3-step23852-tokens51B revision.
Shouldn't these be the exact same model according to https://github.com/allenai/OLMo?

For the 1B model, we have trained three times with different data order on 50B high quality tokens, used last checkpoint of seed 42 as final checkpoint.

yukiwuki

Mar 23

•

edited Mar 23

Hey @aldakata , I am facing the same issue now. Did you manage to figure out what's wrong?

aldakata

Mar 23

I couldn't figure it out, sorry

DavidHidary

Aug 19

Ya I think this is just a bug with the checkpoints, stage2-ingredient3 seems correct.

Nhckdvrl

12 days ago

Weight-level evidence that may explain the mismatch reported here and in #2: main does not come from the same training run as the released stage1/stage2 branches, while -Instruct does descend from stage2-ingredient3.

Summary. The model card says main is the final checkpoint and that fine-tuning can start from it, and that for the 1B model only stage-2 ingredient 3 is the real mid-training run. In the published weights, however, main is unrelated to every stage1-* / stage2-ingredient* branch, including stage2-ingredient3-step23852-tokens51B, while the -Instruct model descends from the stage-2 ingredient-3 branch. main looks like a model from a different training run (different initialization).

Evidence (Pearson correlation of flattened tensors, float64, read from the published safetensors):

comparison	layer-5 q_proj	layer-5 v_proj	layer-5 mlp.down_proj	embed_tokens
`main` vs `stage2-ingredient3-step23852-tokens51B`	0.0003	0.0002	−0.0004	0.0058
`main` vs `stage1-step1907359-tokens4001B`	0.0004	—	—	0.0059
`main` vs `stage1-step0-tokens0B`	0.0001	—	—	0.0024
`stage1-step1907359` vs `stage2-ingredient3-step23852`	0.960	0.977	0.971	0.9996
`OLMo-2-0425-1B-Instruct` vs `main`	—	—	−0.0004	—
`OLMo-2-0425-1B-Instruct` vs `stage2-ingredient3-step23852`	—	—	0.995	—

The configs of main and the stage-2 branch are identical, so this is not a format difference. The near-zero correlation also holds for v_proj, down_proj and the embeddings, which a q/k RoPE permutation in conversion would not affect.
Functionally, main and stage2-ingredient3-step23852 give different next-token distributions (mean KL 0.21 nats/token, top-1 agreement 0.78 on three short prompts), i.e. they are different models, not the same weights stored differently.
stage1-step0-tokens0B vs stage1-step1907359 gives 0.005, as expected for a single run; main is equally far from all branches.

Why it matters. Work that treats main as the endpoint of the released stage1 → stage2 trajectory, or as the base model of -SFT/-DPO/-Instruct (e.g. base-vs-instruct comparisons, training-dynamics analyses that append main to the checkpoint series), compares unrelated models. Could you clarify which run main comes from, or point main to the intended final checkpoint?

Reproduce: read one tensor per revision via huggingface_hub.hf_hub_url("allenai/OLMo-2-0425-1B", <shard>, revision=...) with an HTTP Range request, e.g. model.layers.5.mlp.down_proj.weight, and compute the correlation.

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