Marin-Snowball-67B-A2B-v0

Instructions to use open-athena/Marin-Snowball-67B-A2B-v0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use open-athena/Marin-Snowball-67B-A2B-v0 with Transformers:

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

pipe = pipeline("text-generation", model="open-athena/Marin-Snowball-67B-A2B-v0")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("open-athena/Marin-Snowball-67B-A2B-v0", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use open-athena/Marin-Snowball-67B-A2B-v0 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "open-athena/Marin-Snowball-67B-A2B-v0"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "open-athena/Marin-Snowball-67B-A2B-v0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/open-athena/Marin-Snowball-67B-A2B-v0

SGLang

How to use open-athena/Marin-Snowball-67B-A2B-v0 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 "open-athena/Marin-Snowball-67B-A2B-v0" \
    --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": "open-athena/Marin-Snowball-67B-A2B-v0",
		"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 "open-athena/Marin-Snowball-67B-A2B-v0" \
        --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": "open-athena/Marin-Snowball-67B-A2B-v0",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use open-athena/Marin-Snowball-67B-A2B-v0 with Docker Model Runner:
```
docker model run hf.co/open-athena/Marin-Snowball-67B-A2B-v0
```

Marin Snowball 67B-A2B v0

Research artifact. This model has not been properly tested or evaluated and is not necessarily secure. Use at your own risk in production settings.

This is the dare-linear-02 checkpoint from the third Snowball/Grug model-merging campaign. It merges the allkimi relay SWE/Terminal specialist into the Campaign 2 DARE-TIES-02 base: DARE-linear with density 0.20, delta scale 0.25 and seed 0 within a fixed routed-expert support of 12.4860469471% of Base scalars. After BF16 storage, 1,581,121,835 of 67,078,882,816 scalar parameters differ from Base (2.3571081816%); 97.6428918184% of Base scalars remain bit-for-bit unchanged. All non-weight model files, tensor names, shapes, dtypes, shard layout and index are unchanged from Base.

Fixed-policy benchmark	Score (%)	Scored / declared trials
SWE-Bench-100, mini-swe-agent 2.1.0	33.00	100 / 100
Terminal-Bench 2.1, Terminus 2	13.36	262 / 267
Sum	46.36

Base scored 37.20 (27.27 + 9.92) and the Specialist 24.36 (9.09 + 15.27) under the same policy; the serial-01 candidate reached 48.17. These are development benchmark results, not held-out evidence. Non-target capability retention was not evaluated for this checkpoint.

Parent checkpoints are pinned as follows:

Base: open-athena/Grug-67B-A2B-Merge-Campaign2-DARE-TIES-02-2026.10.10 at 153f5369f69a0c0a7203779099ec94e98c66748f.
Specialist: laion/snowball-67b-a2b-relay-sft-allkimi-step1203 at ff664b48f4c79a81c585dbc9e082b34af3a22b4a.

The complete construction record is preserved in provenance/campaign3/: the merge audit with source hashes, parent pins, and independently verified SWE-Bench-100 and TB2.1 aggregates. The campaign's recipe, protocol and complete 32-candidate tracker are public. Root merge provenance summarizes the same construction. Source S3 paths inside those records are historical storage identifiers, not public download links.

The inherited .cache_metadata.json and .marinskyrl-model-manifest.json describe the Base checkpoint's original cache identity and remain unchanged for source fidelity; use provenance/campaign3/audit.json to verify this merged checkpoint's files.

Publication correction (2026-10-12)

The original publication preserved all 51 source files byte-for-byte, including the root model card and merge-provenance.json, which described Campaign 2 dare-ties-02. This card and the root provenance replace those inherited files with the Campaign 3 dare-linear-02 identity. No weight file was changed by this correction; the weights remain those published at revision 271c4b3c6b38e8f645c88b8938a6ccf2b2195f2a.

Licensed under OpenMDW 1.1.

We release our tools to the public with the intent that they are used safely, ethically, and responsibly. They are meant to contribute to the public's understanding of science, and aid in its advancement. By using this Marin model, you agree to comply with all applicable law.

Downloads last month: 16

Safetensors

Model size

67B params

Tensor type

F32

BF16

Model tree for open-athena/Marin-Snowball-67B-A2B-v0

laion/snowball-67b-a2b-relay-sft-allkimi-step1203

open-athena/Grug-67B-A2B-Merge-Campaign2-DARE-TIES-02-2026.10.10

Merge model

this model

Collection including open-athena/Marin-Snowball-67B-A2B-v0

Releases

Collection

2 items • Updated 1 day ago