JUWEL Emerald
JUWEL Emerald (internal lineage: Vext-Base-v9) — the efficient, general-purpose Vext Labs base. Emerald is the May birthstone: evergreen. ~32B, vision-capable, the base most of the GEM specialist adapters are trained against. Released under Apache-2.0 as part of the JUWEL open archive.
At a glance
| Field | Value |
|---|---|
| Public name | JUWEL Emerald |
| HF repo | VextLabsinc/juwel-emerald |
| Internal lineage | Theron-Base v9 (Qwen3-VL-32B + CIP additive) |
| Parameters | ~32B |
| Precision | BF16 (no quantization) |
| Source | 14 safetensors shards, ~67 GB |
| Release license | Apache-2.0 |
| Hardware floor | 1x H100 80GB (BF16) |
What this is, honestly
JUWEL Emerald is the dense, efficient general base in the Vext lineage — Qwen3-VL-32B with a CIP additive step. It is the base the GEM specialist fleet loads on top of. It is released as open weights because it is a strong, useful base and because the community feedback loop on it feeds how we build forward.
Intended use
- General reasoning / language / vision baseline at ~32B.
- Base for the GEM specialist adapters (
VextLabsinc/gem-*) and for your own domain fine-tunes. - Self-hosted deployment without depending on Vext-hosted inference.
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 human-to-model boundary.
- Quantized inference if quality is the goal — trained/validated in BF16.
Architecture / training (outcome-level)
- Derived from Qwen3-VL-32B; capability added via CIP (organic upscale + LoRA on new layers, then merged). We describe CIP at the outcome level only; the recipe is proprietary.
- No teacher distillation — training examples are curated primary sources. Raw corpus stays proprietary; a data card describing source families ships with the release.
Evaluation
- Status: pending — numbers published only when reproducible on our stack (named harness, date, method), per our benchmark-audit-trail discipline. Do not treat internal rubrics as public leaderboards. Headline lineage note: v9 beat the retired v8 by a wide margin on GSM8K / MMLU-Pro at one-third the footprint (audit trail: https://vextlabs.ai/transparency).
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-emerald
mkdir -p juwel-emerald && cd juwel-emerald
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 14); do curl -O $BASE/model-000$i-of-00014.safetensors; done
How to load
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained("./juwel-emerald", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("./juwel-emerald")
Load a GEM specialist on top (download its adapter from R2 first — see the GEM's card — then):
from peft import PeftModel
model = PeftModel.from_pretrained(model, "./gem-ruby") # code specialist, downloaded from R2
License / attribution
Derived from Qwen3-VL-32B (Alibaba Cloud). Preserve the base license + NOTICE. JUWEL Emerald
is released under Apache-2.0. Do not imply Alibaba/Qwen endorsement. Contact: info@vextlabs.ai.