We expand on our Olmo model series by introducing Olmo Hybrid, a new 7B hybrid RNN model in the Olmo family. Olmo Hybrid dramatically outperforms Olmo 3 in final performance, consistently showing roughly 2x data efficiency on core evals over the course of our pretraining run. We also show gains in performance on long-context benchmarks, as well as improved inference efficiency (throughput and memory) on long-context lengths by a factor of 75%.

The core models released in this batch include the following:

Stage	Olmo 3 7B Think	Olmo 3 32B Think	Olmo 3 7B Instruct	Olmo Hybrid Think 7B	Olmo Hybrid Instruct 7B
Base Model	Olmo-3-7B	Olmo-3-32B	Olmo-3-7B	Olmo-Hybrid-7B	Olmo-Hybrid-7B
SFT	Olmo-3-7B-Think-SFT	Olmo-3-32B-Think-SFT	Olmo-3-7B-Instruct-SFT	Olmo-Hybrid-Think-SFT-7B	Olmo-Hybrid-Instruct-SFT-7B
DPO	Olmo-3-7B-Think-DPO	Olmo-3-32B-Think-DPO	Olmo-3-7B-Instruct-DPO	--	Olmo-Hybrid-Instruct-DPO-7B
Final Models (RLVR)	Olmo-3-7B-Think	Olmo-3-32B-Think	Olmo-3-7B-Instruct	--	--

Olmo is a series of Open language models designed to enable the science of language models. These models are pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets. We are releasing all code, checkpoints, logs (coming soon), and associated training details.

Installation

Olmo Hybrid is supported in transformers 5.3.0 or higher

pip install transformers>=5.3.0

Inference

You can use OLMo with the standard HuggingFace transformers library:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
olmo = AutoModelForCausalLM.from_pretrained("allenai/Olmo-Hybrid-Think-SFT-7B")
tokenizer = AutoTokenizer.from_pretrained("allenai/Olmo-Hybrid-Think-SFT-7B")
message = ["Who would win in a fight - a dinosaur or a cow named Moo Moo?"]
inputs = tokenizer(message, return_tensors='pt', return_token_type_ids=False)
# optional verifying cuda
# inputs = {k: v.to('cuda') for k,v in inputs.items()}
# olmo = olmo.to('cuda')
response = olmo.generate(**inputs, max_new_tokens=100, do_sample=True, top_k=50, top_p=0.95)
print(tokenizer.batch_decode(response, skip_special_tokens=True)[0])
>> '<think>Okay, so the question is who would win in a fight...'

For faster performance, you can quantize the model using the following method:

python

AutoModelForCausalLM.from_pretrained("allenai/Olmo-Hybrid-Think-SFT-7B", 
    torch_dtype=torch.float16, 
    load_in_8bit=True)  # Requires bitsandbytes

The quantized model is more sensitive to data types and CUDA operations. To avoid potential issues, it's recommended to pass the inputs directly to CUDA using:

inputs.input_ids.to('cuda')

We have released checkpoints for these models. For post-training, the naming convention is step_XXXX.

To load a specific model revision with HuggingFace, simply add the argument revision:

olmo = AutoModelForCausalLM.from_pretrained("allenai/Olmo-Hybrid-Think-SFT-7B", revision="step11000")

Or, you can access all the revisions for the models via the following code snippet:

python

from huggingface_hub import list_repo_refs
out = list_repo_refs("allenai/Olmo-Hybrid-Think-SFT-7B")
branches = [b.name for b in out.branches]

Chat template

Default System Message

The default system prompt for this model is:

<|im_start|>system
You are a helpful function-calling AI assistant. You do not currently have access to any functions. <functions></functions>
<|im_end|>

Chat Format

The chat template for this model is formatted as:

<|im_start|>system
You are a helpful function-calling AI assistant. You do not currently have access to any functions. <functions></functions>
<|im_start|>user
Who would win in a fight - a dinosaur or a cow named Moo Moo?<|im_end|>
<|im_start|>assistant
<think>Okay, so the question is who would win in a fight between a dinosaur and a cow named Moo Moo.
Hmm, first I need to break this down. Let me think about the different factors involved here..... </think>
Moo Moo the cow would certinaly win.
<|endoftext|>

Model Description

Developed by: Allen Institute for AI (Ai2)
Model type: a Transformer style autoregressive language model.
Language(s) (NLP): English
License: This model is licensed under Apache 2.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.
Contact: Technical inquiries: olmo@allenai.org. Press: press@allenai.org
Date cutoff: Dec. 2024.

Model Sources

Project Page: https://allenai.org/olmo
Repositories:
- Open-Instruct for DPO and RLVR: https://github.com/allenai/open-instruct
- OLMo-Core for pre-training and SFT: https://github.com/allenai/OLMo-core
- OLMo-Eval for evaluation: https://github.com/allenai/OLMo-Eval
Olmo 3 Paper: https://allenai.org/papers/olmo3
Olmo Hybrid Paper: https://allenai.org/papers/olmo-hybrid

Evaluation

Skill	Benchmark	Olmo Hybrid Think SFT 7B	Olmo 3 Think 7B SFT	Olmo 3 Think 7B DPO	Olmo 3 Think 7B	OpenThinker3-7B	Nemotron-Nano-9B-v2	DeepSeek-R1-Distill-Qwen-7B	Qwen 3 8B (reasoning)	Qwen 3 VL 8B Thinker	OpenReasoning Nemotron 7B
Math	MATH	93.8	94.4	92.4	95.1	94.5	94.4	87.9	95.1	95.2	94.6
	AIME 2024	66.2	69.6	74.6	71.6	67.7	72.1	54.9	74.0	70.9	77.0
	AIME 2025	55.2	57.6	62.7	64.6	57.2	58.9	40.2	67.8	61.5	73.1
	OMEGA	35.1	45.0	40.5	37.8	38.4	42.4	28.5	43.4	38.1	43.2
Reasoning	BBH	84.6	84.1	83.7	86.6	77.1	86.2	73.5	84.4	86.8	81.3
	ZebraLogic	55.1	57.9	60.6	66.5	34.9	60.8	26.1	85.2	91.2	22.4
	AGI Eval		77.2	79.1	81.5	78.6	83.1	69.5	87.0	90.1	81.4
Coding	HumanEval+	86.3	88.2	91.4	89.9	87.4	89.7	83.0	80.2	83.7	89.7
	MBPP+	63.7	63.2	63.0	64.7	61.4	66.1	63.5	69.1	63.0	61.2
	LCB v3	65.5	67.8	75.1	75.2	68.0	83.4	58.8	86.2	85.5	82.3
IF	IFEval	80.4	77.9	75.9	88.2	51.7	86.0	59.6	87.4	85.5	42.5
	IFBench	31.6	30.0	28.3	41.6	23.0	34.6	16.7	37.1	40.4	23.4
Knowledge	MMLU	80.5	74.9	74.8	77.8	77.4	84.3	67.9	85.4	86.5	80.7
QA	PopQA	25.1	20.8	24.7	23.7	18.0	17.9	12.8	24.3	29.3	14.5
	GPQA	47.0	45.8	48.6	46.2	47.6	56.2	54.4	57.7	61.5	56.6
Chat	AE 2	49.0	43.9	50.6	52.1	24.0	58.0	7.7	60.5	73.5	8.6
Safety			65.8	67.7	70.7	31.3	72.1	54.0	68.3	82.9	30.3

Model Details

SFT

supervised fine-tuning on the Dolci-Think-SFT-7B dataset. This dataset consits of math, code, chat, and general knowledge queries.
Datasets: Dolci-Think-SFT-7B

Inference & Recommended Settings

We evaluated our models on the following settings. We also recommend using them for generation:

temperature: 0.6
top_p: 0.95
max_tokens: 32768

ONNXRuntime

from transformers import AutoConfig, AutoTokenizer
import onnxruntime
import numpy as np
from huggingface_hub import snapshot_download

# 1. Load config, processor, and model
model_id = "onnx-community/Olmo-Hybrid-Think-SFT-7B-ONNX"
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

model = "model_q4f16.onnx"
model_folder = snapshot_download(model_id, allow_patterns=[f"onnx/{model}*"])
session = onnxruntime.InferenceSession(f"{model_folder}/onnx/{model}")

## Set config values
num_key_value_heads = config.num_key_value_heads
head_dim = config.hidden_size // config.num_attention_heads
num_hidden_layers = config.num_hidden_layers
eos_token_id = config.eos_token_id

# 2. Prepare inputs
prompt = "Who would win in a fight - a dinosaur or a cow named Moo Moo?"
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="np")
input_ids = inputs['input_ids']
attention_mask = inputs['attention_mask']
batch_size = input_ids.shape[0]
cache = {}
for inp in session.get_inputs():
  name = inp.name
  dtype = np.float16 if inp.type == "tensor(float16)" else np.float32
  if name.startswith("past_key_values"):
    cache[name] = np.zeros([batch_size, num_key_value_heads, 0, head_dim], dtype=dtype)
  elif name.startswith("past_conv") or name.startswith("past_recurrent"):
    shape = [batch_size if i == 0 else dim for i, dim in enumerate(inp.shape)]
    assert all(isinstance(dim, int) for dim in shape[1:]), f"Unexpected dynamic cache shape: {inp.shape}"
    cache[name] = np.zeros(shape, dtype=dtype)

# 3. Generation loop
max_new_tokens = 1024
generated_tokens = np.array([[]], dtype=np.int64)
for i in range(max_new_tokens):
  logits, *present_cache_values = session.run(None, dict(
    input_ids=input_ids,
    attention_mask=attention_mask,
    **cache,
  ))

  ## Update values for next generation loop
  input_ids = logits[:, -1].argmax(-1, keepdims=True)
  attention_mask = np.concatenate([attention_mask, np.ones_like(input_ids, dtype=np.int64)], axis=-1)
  for j, key in enumerate(cache):
    cache[key] = present_cache_values[j]
  generated_tokens = np.concatenate([generated_tokens, input_ids], axis=-1)
  if np.isin(input_ids, eos_token_id).any():
    break

  ## (Optional) Streaming
  print(tokenizer.decode(input_ids[0]), end='', flush=True)
print()

# 4. Output result
print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0])

Bias, Risks, and Limitations

Like any base language model or fine-tuned model without safety filtering, these models can easily be prompted by users to generate harmful and sensitive content. Such content may also be produced unintentionally, especially in cases involving bias, so we recommend that users consider the risks when applying this technology. Additionally, many statements from OLMo or any LLM are often inaccurate, so facts should be verified.

License

This model is licensed under Apache 2.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Citation

Coming Soon!

Model Card Contact

For errors in this model card, contact olmo@allenai.org.