JUWEL Sapphire
JUWEL Sapphire (internal lineage: Vext-Base-v10, iter-4) — the vision-native Vext Labs base. Sapphire is the September birthstone: clarity and sight. ~72B, 144 layers, organically upscaled across four CIP iterations from Emerald. This is the base the padded GEM fleet targets (144L). Released under Apache-2.0 as part of the JUWEL open archive.
At a glance
| Field | Value |
|---|---|
| Public name | JUWEL Sapphire |
| HF repo | VextLabsinc/juwel-sapphire |
| Internal lineage | Theron-Base v10 iter-4 (Qwen3-VL-derived, organic upscale) |
| Parameters | ~72B, 144 layers, hidden 5120 |
| Context | 262K |
| Precision | BF16 (no quantization) |
| Source | 30 safetensors shards, ~145 GB |
| Release license | Apache-2.0 |
| Hardware floor | 2x H100 80GB (BF16); 4x recommended |
What this is, honestly
JUWEL Sapphire is the vision-native 72B base — DeepStack vision indexing, 144 layers reached by organic upscale (16 new layers per CIP iteration) over the Emerald lineage. It is the recommended base for the 144-layer GEM adapters. Released as open weights so the community can build on the current-generation-behind flagship.
Intended use
- Vision-language + reasoning at ~72B.
- Base for the 144-layer GEM specialist adapters (
VextLabsinc/gem-*, padded to 144L). - Self-hosted deployment; research on organic-upscale / composition architectures.
Out of scope
- Not a substitute for a licensed professional in any regulated domain.
- No deliberate refusal training — apply your own safety policy at the boundary.
- Quantized inference if quality is the goal — trained/validated in BF16 (third parties may quant).
Architecture / training (outcome-level)
- Qwen3-VL-derived; grown via CIP (organic upscale +16 layers/iter, LoRA on new layers, merged). CIP is described at the outcome level only; the recipe is proprietary.
- No teacher distillation — curated primary sources; raw corpus proprietary, data card included.
Evaluation
- Status: pending — published only when reproducible on our stack (named harness, date, method). No fabricated scores. Internal rubrics are not public leaderboards.
Weights — download (Cloudflare R2, public, BF16)
Weights live on Cloudflare R2, not inside this HF repo. Download them locally, then load from the local dir.
BASE=https://pub-a6ae0476e46849f98f1746a61dc4c106.r2.dev/juwel-sapphire
mkdir -p juwel-sapphire && cd juwel-sapphire
for f in config.json model.safetensors.index.json tokenizer.json tokenizer_config.json generation_config.json; do curl -sO $BASE/$f; done
for i in $(seq -w 1 30); do curl -O $BASE/model-000$i-of-00030.safetensors; done
How to load
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained("./juwel-sapphire", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("./juwel-sapphire")
Load a 144L GEM specialist on top (download its adapter from R2 first, then):
from peft import PeftModel
model = PeftModel.from_pretrained(model, "./gem-spinel") # reasoning specialist, downloaded from R2
License / attribution
Derived from the Qwen3-VL lineage (Alibaba Cloud). Preserve the base license + NOTICE.
JUWEL Sapphire is released under Apache-2.0. Do not imply Alibaba/Qwen endorsement.
Contact: info@vextlabs.ai.