saggeseluca/o1.c-models

HiDream-O1-Image — GGUF conversions for the native o1.c runtime

This repository hosts GGUF conversions of the HiDream-O1-Image model family, packaged for the native o1.c inference engine (C/CUDA, no PyTorch required at runtime).

Runtime: github.com/luca-saggese/o1.c — clone, build and run the engine to use these files.

These are GGUF conversions for the native o1.c runtime.

Original model authors: HiDream-ai. All model weights, architecture and training originate from the upstream HiDream-ai repositories. This repository does not replace the upstream model repository; it only redistributes a reformatted (GGUF) copy of the released weights under the same MIT license.

If you want the original PyTorch/safetensors checkpoints, or the authoritative model card, training details and citation, please use the upstream repositories listed under Provenance.

---

Available artifacts

All artifacts are BF16 GGUF files with the o1.c layout version 1.

File	Profile	Variant	Steps	Size	SHA256
hidream-o1-dev-2604-bf16.gguf	dev	dev-2604	28	17 609 841 152 B	7c405035a9225b721ba19b1a1f4e7f1cb937ddea7dbb7207e5ad677c3e0a7f5d
hidream-o1-dev-bf16.gguf	dev	dev	28	17 609 841 152 B	4b1141733ec5a99fcbc2994bf720e8ef6ade31ec3e9378e471e73f3af6ce8fae
hidream-o1-base-bf16.gguf	base	base	50	17 609 841 152 B	07f5efb347cfd9264c8dadd3848bb8c6053633b37a05b0680d920a6409de011b

hidream-o1-dev-2604-bf16.gguf is the recommended artifact.

The checksums are also published in SHA256SUMS and in the o1.c engine repository's models/manifest.json.

Q4 is not available. The o1.c runtime currently accepts only F32/F16/BF16 GGUF tensors; there is no quantized weight path yet. Do not expect or request Q4 files from this repository until a separate quantization gate is completed upstream in o1.c.

---

Provenance

Each artifact is a byte-exact re-layout of one immutable upstream revision. The revision is a commit SHA, never a branch name.

Artifact	Source repository	Source revision	License
hidream-o1-dev-2604-bf16.gguf	HiDream-ai/HiDream-O1-Image-Dev-2604	b6acc2fe452b3120430620dc4354fa442ee081ea	MIT
hidream-o1-dev-bf16.gguf	HiDream-ai/HiDream-O1-Image-Dev	c0bada0e15c54a9f96a6d1ecc35575b32bc21544	MIT
hidream-o1-base-bf16.gguf	HiDream-ai/HiDream-O1-Image	0b0901d99f200389e138c61946af1185f5f49a13	MIT

Upstream license: MIT (declared in the model metadata of all three source repositories). The upstream repositories do not ship a standalone LICENSE file, so the MIT text is reproduced in LICENSE for convenience and no additional terms are imposed by this repository.

No weights were modified, fine-tuned, merged, pruned or re-quantized. The conversion is a layout/dtype-preserving re-serialization: tensor values are the upstream values in the upstream dtype (BF16), stored in the o1.c GGUF container with its metadata keys added.

---

GGUF metadata

Every artifact carries at least the following keys, which the o1.c runtime reads to configure itself (so that a single --model-path is enough):

general.architecture        = hidream_o1
general.name                = <model name>
hidream.profile             = dev | base
hidream.variant             = dev | dev-2604 | base
hidream.revision            = <upstream commit sha>
hidream.dtype               = bf16
hidream.quantization        = bf16
hidream.num_layers          = <int>
hidream.layout_version      = 1

Execution profile derived from hidream.profile:

• dev → 28 steps, guidance 0, shift 1
• base → 50 steps, guidance 5, shift 3

---

How to use

These files are only usable with the o1.c runtime — a native C/CUDA inference engine that loads GGUF weights directly, with no PyTorch or Python at runtime.

Engine repository: https://github.com/luca-saggese/o1.c

Clone the engine, build it, then download and run a model:

# 1. get the engine
git clone https://github.com/luca-saggese/o1.c.git
cd o1.c

# 2. check your GPU/CUDA/cuDNN toolchain and build
./scripts/setup.sh
make

# 3. download a model from this repository (checksum-verified)
./scripts/download_model.sh dev-2604

# 4. generate an image
./build/hidream --model-path models/hidream-o1-dev-2604-bf16.gguf \
    --prompt "a red fox in a snowy forest, golden hour" \
    --width 2048 --height 2048 --seed 42 --output fox.png

The engine README covers installation, the full CLI reference, image editing, multi-reference personalization, skeleton/layout conditioning, the OpenAI-compatible server and troubleshooting.

You can also download the GGUF files from this repository directly and pass the local path to --model-path.

---

Acknowledgements

Model architecture, training and original weights: HiDream-ai. GGUF conversion and the native runtime: the o1.c project — these files require that engine to run.

If you use these weights, please cite the original HiDream-O1-Image work as described in the upstream model cards.