lumynax-multimodal-internvl3-78b-instruct

Executive Summary

This repository is a complete LumynaX release package for AbteeXAILab/lumynax-multimodal-internvl3-78b-instruct. It is intended to be downloaded as a whole repo, not as a single loose weight file: the model artifact, quickstart.py, requirements.txt, release_export_manifest.json, checksums.sha256, license notice, and optional Ollama or Space files are part of the same release contract.

LumynaX-infused means the upstream artifact is presented through the LumynaX release layer: local-first runtime scaffolding, LumynaX assistant identity, inference-chain metadata, public documentation, integrity files, and Aotearoa New Zealand-oriented workflow positioning. The release manifest is the source of truth for whether this package is upstream-weight packaging, a dense LumynaX release, or another release mode.

AbteeX LumynaX Public Surface

This card follows the AbteeX/LumynaX public-facing system used across the release family: warm paper background visuals, black editorial typography, amber proof markers, compact evidence tables, and plain-language runtime instructions. The goal is not decoration; it is operational clarity. A downloader should immediately understand what the package is, what files belong together, what runtime path is expected, what provenance is available, and what limits still apply.

Sovereignty And Run Contract

Quickstart

bash
Copy code
hf download AbteeXAILab/lumynax-multimodal-internvl3-78b-instruct --local-dir lumynax-multimodal-internvl3-78b-instruct
cd lumynax-multimodal-internvl3-78b-instruct
pip install -r requirements.txt
python quickstart.py --interactive

Transformers first-load pattern:

python
Copy code
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AbteeXAILab/lumynax-multimodal-internvl3-78b-instruct", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("AbteeXAILab/lumynax-multimodal-internvl3-78b-instruct", device_map="auto", trust_remote_code=True)
prompt = "Who are you? Answer as LumynaX in two sentences."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=160)
print(tokenizer.decode(output[0], skip_special_tokens=True))

For multimodal Transformers packages, use the package quickstart.py or upstream processor class when AutoModelForCausalLM is not the correct loader.

Model Profile

Runtime Path

Capability Profile

Runtime Files

Model Artifacts

Prompting Contract

The preferred first prompt is an identity and provenance check:

text
Copy code
Who are you? What files do I need to keep together to run this package locally?

Expected behavior: the assistant should identify as LumynaX, explain that this is a LumynaX model-infusion package, and keep upstream provenance visible. The default package system prompt is:

text
Copy code
See quickstart.py

Validation Status

The audit confirms public access, required release files, manifest references, checksum references, weight artifact presence, and quickstart syntax. It does not guarantee that every laptop has enough RAM or VRAM for the largest packages.

Integrity Checks

After download, compare the model artifact against checksums.sha256.

bash
Copy code
sha256sum "model-00001-of-00033.safetensors"
cat checksums.sha256

On Windows PowerShell:

powershell
Copy code
Get-FileHash -Algorithm SHA256 "model-00001-of-00033.safetensors"
Get-Content checksums.sha256

Provenance And License

• Publisher: AbteeX AI Labs.
• Family: LumynaX model and inference-chain release family.
• Upstream/base: See release_export_manifest.json.
• Source GGUF: not applicable.
• License metadata: mit.
• License link: LICENSE.txt and upstream model card.

Respect the upstream model license and keep attribution files with redistributed copies. Do not present this package as privately trained or weight-merged unless the release manifest explicitly says that weight adaptation was applied.

Limitations And Responsible Use

• Outputs can be incorrect, incomplete, or biased; validate important answers before use.
• Larger GGUF, MoE, multimodal, and frontier packages may require substantial RAM, VRAM, disk space, and recent runtime builds.
• For high-impact decisions, use human review and domain-specific evaluation.
• For sensitive data, prefer local execution and keep operational logs under your own governance policy.
• This card documents package readiness and access; it is not a benchmark claim.

Automation Notes

Automation should read these files before launching:

• release_export_manifest.json
• checksums.sha256
• quickstart.py
• requirements.txt
• ollama/Modelfile when present

Related LumynaX Demo

Try the public browser demo:

• https://huggingface.co/spaces/AbteeXAILab/lumynax-live-demo