MolmoPoint-8B-bnb-int8-native

MolmoPoint-8B — bitsandbytes LLM.int8

This is an inference-ready bitsandbytes LLM.int8 quantization of allenai/MolmoPoint-8B. All eligible linear modules are quantized except the 7.5M-parameter pointing head, which remains BF16. This avoids a runtime patch because bitsandbytes 0.50.0's INT8 kernel does not natively accept the pointing head's high-rank inputs.

Evaluation overview

The figure compares both released quantizations with the BF16 baseline. See Evaluation for coverage, protocol, and metric details.

Model details

• Repository: erdi28/MolmoPoint-8B-bnb-int8-native
• Base model: allenai/MolmoPoint-8B
• Base revision: 188130f961c8e0888a34e11121a1423c461a01ba
• Quantized: all eligible linear modules outside model.point_predictor
• Retained in BF16: model.point_predictor (7.5M parameters; under 0.1% of the model)
• Outlier threshold: 6.0
• Local checkpoint size: approximately 9.3 GiB
• Source and evaluation code: erkara/molmopoint-quantization

The included quantization_provenance.json records the resolved base revision, complete quantization recipe, software versions, and build environment.

Requirements

The checkpoint was built and validated with Python 3.12, Transformers 4.57.1, bitsandbytes 0.50.0, and a CUDA-capable NVIDIA GPU. It contains the processor and trusted custom model code required for native Transformers loading.

pip install "transformers==4.57.1" "bitsandbytes==0.50.0" accelerate torch torchvision pillow einops decord2

Usage

import torch
from PIL import Image
from huggingface_hub import hf_hub_download
from transformers import AutoModelForImageTextToText, AutoProcessor

repo_id = "erdi28/MolmoPoint-8B-bnb-int8-native"

processor = AutoProcessor.from_pretrained(
    repo_id,
    trust_remote_code=True,
    padding_side="left",
    use_fast=False,
)
model = AutoModelForImageTextToText.from_pretrained(
    repo_id,
    trust_remote_code=True,
    dtype="auto",
    device_map="auto",
)
model.eval()

image_path = hf_hub_download(repo_id, "examples/pointing-demo.jpg")
image = Image.open(image_path).convert("RGB")
prompt = "Point to the tool that people can use to write."
messages = [{
    "role": "user",
    "content": [
        {"type": "text", "text": prompt},
        {"type": "image", "image": image},
    ],
}]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    padding=True,
    return_pointing_metadata=True,
)
metadata = inputs.pop("metadata")
inputs = {name: value.to("cuda") for name, value in inputs.items()}
prompt_length = inputs["input_ids"].shape[1]

with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
    output = model.generate(
        **inputs,
        logits_processor=model.build_logit_processor_from_inputs(inputs),
        do_sample=False,
        max_new_tokens=128,
    )

generated = output[:, prompt_length:]
generated_text = processor.post_process_image_text_to_text(
    generated,
    skip_special_tokens=False,
    clean_up_tokenization_spaces=False,
)[0]
points = model.extract_image_points(
    generated_text,
    metadata["token_pooling"],
    metadata["subpatch_mapping"],
    metadata["image_sizes"],
)

print(generated_text)
print(points)

Example output

This checkpoint returns two points near (324, 100) and (627, 187) in the 1200 x 800 source image, corresponding to the blue and red mechanical pencils. The demo image is a compressed copy of a CC0 Public Domain image from PxHere. MolmoPoint emits special grounding tokens, which extract_image_points converts into source-image coordinates.

Evaluation

All results use greedy decoding, batch size 1, the uber_model_v2 prompt templates, and AllenAI's official Molmo2 formatter and evaluators pinned to commit f3cb1085fbb97c4a4d7fdcadd77cf871bf37a88a.

All selected examples passed with zero evaluation failures. PointBench is a complete public evaluation. PixMo-Points publishes 436 metadata rows through allenai/pixmo-points-eval, but only 231 external image URLs still returned bytes matching the published SHA-256 hashes. The PixMo score therefore has 52.98% public-data coverage and is not directly comparable with the paper's full-dataset result.

Peak allocated VRAM was 36.0% lower than BF16 on the benchmark system. INT8 was slower than BF16 in both evaluations; its benefit here is reduced memory and storage rather than speed. Memory and latency are environment-specific and should not be treated as universal hardware requirements. Full category scores and protocol details are available in the project results.

Intended use

This checkpoint is intended for research, evaluation, and CUDA inference where reduced memory use is useful and MolmoPoint's image, multi-image, video, and grounding capabilities are required. It is a quantized derivative, not a fine-tuned model, and does not add new capabilities or safety training.

Limitations

This release inherits the base model's limitations, biases, and failure modes. Quantization can change individual generations even when aggregate scores are close. The evaluation does not establish identical or lossless behavior, and PixMo-Points coverage is partial. Loading requires trust_remote_code=True and a CUDA environment supported by bitsandbytes.

License and responsible use

This derivative retains the base model's Apache-2.0 license and attribution. Use is also subject to Ai2's Responsible Use Guidelines. The upstream model card states that MolmoPoint-8B was trained on third-party datasets subject to academic and non-commercial research-use terms. Review the base model card and applicable source-dataset terms before use or redistribution.