JUWEL Sapphire
Release status: MIRRORED_STRUCTURAL_PASS. Every mirrored object matched the pinned public-R2 inventory, was SHA-256 hashed, and passed Hugging Face readback. Safetensors shards also matched the published index at header/tensor/offset level.
This repository contains public model weights mirrored from Vext Labs' already-public Cloudflare
R2 release at https://pub-a6ae0476e46849f98f1746a61dc4c106.r2.dev/juwel-sapphire. It is not JUWEL's own-weights flagship Theta, and this
card makes no SOTA, production-safety, or benchmark claim.
Artifact-derived specification
- Architecture:
Qwen3VLForConditionalGeneration - Geometry: 72,365,256,944 parameters; BF16; 144 text layers; hidden size 5,120; FFN 25,600; 64 attention heads / 8 KV heads; 262,144-token configured context; vocabulary 151,936; 27-layer vision tower, hidden size 1,152, patch size 16
- Package: 30 safetensors shards; 1,938 indexed tensors; 144,730,513,888 tensor bytes
- Lineage: Qwen3-VL architecture family. The recovered release metadata does not bind an exact upstream revision; no endorsement by Qwen is implied.
The exact source inventory and file hashes are recorded in
mirror-receipts/4a500af65230058831c68cd323b836d3305e8dbbc3257d9f6b8e1d2cdbd42a2a.json.
Loading
from transformers import AutoModelForImageTextToText, AutoProcessor
repo = "VextLabsinc/juwel-sapphire"
model = AutoModelForImageTextToText.from_pretrained(repo, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(repo)
These are large BF16 artifacts. Use hardware and sharding appropriate to the package size.
Verification and limitations
The release receipt proves source inventory binding, full-file SHA-256, safetensors structural consistency, and destination byte readback. It does not prove full GPU inference, output quality, training-data provenance, safety behavior, or production fitness. Capability evaluation is pending. Review the exact configuration and receipt before use.
License
Apache License 2.0. The complete license text is in LICENSE. “Qwen” describes architecture
lineage only and does not imply upstream endorsement. This repository includes no additional
field-of-use restriction; Apache-2.0 governs the mirrored files. This is an Apache-licensed
open-weight release, not a claim that the package satisfies the OSI Open Source AI Definition,
which also considers training-data information and modification materials.