AbteeXAILab / lumynax-multimodal-glm46v-flash

LumynaX Multimodal GLM 4.6V Flash

LumynaX Multimodal GLM 4.6V Flash is a LumynaX model-infusion release from AbteeX AI Labs. It is packaged for public-facing Hugging Face distribution with explicit upstream provenance, local runtime instructions, checksums, and LumynaX release metadata.

LumynaX Infusion Identity

LumynaX is the AbteeX AI Labs model and inference-chain family oriented around Aotearoa New Zealand data, local context, and practical edge deployment. In this release, the upstream model artifact is preserved and fed through the LumynaX release layer: packaging, prompts, inference-chain metadata, documentation, local runners, and validation records.

This means the package should be understood as a LumynaX infusion release: you are running the listed upstream model artifact through LumynaX's NZ-data-oriented inference workflow and release identity. It does not remove the upstream license or provenance, and it does not claim a private weight merge unless the manifest explicitly says so.

Release Summary

Download And Run

# Download the complete runnable package.
hf download AbteeXAILab/lumynax-multimodal-glm46v-flash --local-dir lumynax-multimodal-glm46v-flash

# Text smoke through llama.cpp. For image use, add: --image path\to\image.jpg
py -3 lumynax-multimodal-glm46v-flash\quickstart.py --llama-cli path\to\llama-cli.exe --prompt "Reply OK." --max-new-tokens 16

# Direct llama.cpp form:
llama-cli -m lumynax-multimodal-glm46v-flash\lumynax-multimodal-glm46v-flash-ud-iq2_m.gguf -mm lumynax-multimodal-glm46v-flash\mmproj-lumynax-multimodal-glm46v-flash-f16.gguf -p "Describe this image or answer the text prompt." -n 128

Runtime Notes

• Use a current llama.cpp build for GGUF packages. The local smoke environment used llama.cpp build b8840-9e5647aff on Windows.
• For multimodal GGUF packages, download both the model .gguf and matching mmproj file.
• Very large GGUF files require substantial RAM/VRAM. If loading fails, lower context size first, then move to a larger machine.
• Keep all files from the release folder together: README.md , quickstart.py , checksums.sha256 , manifests, license notice, GGUF files, Ollama files, and HF Space files where present.

Included Artifacts

Integrity

Every final package includes checksums.sha256 where available. Verify after download:

Get-FileHash -Algorithm SHA256 path\to\downloaded.gguf
Get-Content checksums.sha256

The expected package metadata is in release_export_manifest.json or family_bundle_manifest.json .

Provenance And License

• LumynaX release publisher: AbteeX AI Labs.
• Upstream/source model: unsloth/GLM-4.6V-Flash-GGUF .
• Weights remain subject to the upstream model license and notices.
• LumynaX documentation, packaging scripts, runners, and release metadata are provided to make the package easier to inspect, download, and run.

Intended Use

This release is intended for local experimentation, inference-chain evaluation, New Zealand-context application prototyping, and reproducible GGUF/edge deployment workflows. Users are responsible for validating outputs, respecting upstream license terms, and using hardware appropriate for the selected quantization.

LumynaX Infusion Identity

This is a public LumynaX release package from AbteeX AI Labs for local-first inference. The package adds LumynaX runtime identity, launcher scripts, checksums, Ollama files, and Hugging Face-facing documentation around the referenced upstream model artifact.

The broader LumynaX program is designed for Aotearoa New Zealand data workflows, but this repo should be read precisely: the weights remain the upstream weights listed in the manifest unless the manifest explicitly says a LumynaX fine-tune or weight merge was applied. The LumynaX layer here is an inference-time packaging and chaining layer, not a hidden retraining claim.

Download And Run

Clone the full repo so the GGUF, quickstart, checksums, and Ollama files stay together:

git lfs install
git clone https://huggingface.co/AbteeXAILab/<repo>
cd <repo>
pip install -r requirements.txt
python quickstart.py --llama-cli /path/to/llama-cli --prompt "Say hello in one short sentence."

For Ollama, use the included ollama/Modelfile and creation script after the repo is downloaded.

Runtime Components

• README.md : this model card and release instructions
• quickstart.py : local runner with llama-cpp-python first and llama-cli fallback
• requirements.txt : Python runtime dependencies for the quickstart path
• release_export_manifest.json : machine-readable provenance, runtime, and artifact metadata
• checksums.sha256 : integrity hashes for the shipped files
• ollama/ : Ollama packaging files
• hf_space/ : lightweight Hugging Face Space scaffold
• packaged source: see release_export_manifest.json
• upstream base model: see release_export_manifest.json
• supported modalities: see manifest
• license metadata: see the model-card front matter and LICENSE.txt

Smoke Status

Hub package audit pending.

Hardware still matters. Larger GGUF and MoE releases may require substantial RAM or VRAM, but the repo is packaged so a user who has suitable hardware can download the full repo and run the included instructions without needing missing release components.