GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
VextLabsinc
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
VextLabsinc
Model Tree
Qwen/Qwen3-VL-32B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Architecture | Qwen3VLForConditionalGeneration (image-text-to-text) |
| Layers | 80 (parent had 64; 16 added by CIP, parent's 64 frozen) |
| Hidden size | 5120 |
| Attention heads | 64 (head dim 128) · 8 KV heads |
| Intermediate size | 25600 |
| Context window | 262144 tokens (256K), rope_type: default — no extended-rope variant |
| Vision tower | 27 layers, inherited from the parent |
| Vocab | 151936 — Qwen's tokenizer, shipped unmodified |
| Precision | BF16 (verified in the shard headers, not just declared) |
| Size | 76.7 GiB, 19 shards, 1234 tensors |
| Parent | Qwen3-VL-32B-Instruct (Apache-2.0) |
| License | Apache-2.0 — see LICENSE + NOTICE. AS IS, no warranty. |
Internal lineage: theron-base-v9 plus one capability-injection rung. Layers 0–63 are
the parent's frozen weights; layers 64–79 were added and trained.
python
from transformers import AutoModelForImageTextToText, AutoTokenizermodel = AutoModelForImageTextToText.from_pretrained("VextLabsinc/juwel-beryl", torch_dtype="auto", device_map="auto")tok = AutoTokenizer.from_pretrained("VextLabsinc/juwel-beryl")
Use AutoModelForImageTextToText, not AutoModelForCausalLM — this is a
vision-language architecture and the causal-LM class will not load it.
Footprint: ~82 GB in BF16 (2x 48 GB GPUs or better), ~41 GB at FP8, ~21 GB at NF4.
Beryl is the base for 16 domain LoRA adapters (VextLabsinc/gem-*: ruby/code,
aquamarine/math, spinel/reasoning, lapis/language, bloodstone/medical,
sardonyx/legal, tanzanite/finance, peridot/science, topaz/engineering,
citrine/business, opal/education, jade/humanities, pearl/reconciler, plus
alexandrite, tourmaline, turquoise). Each is native to this exact 80-layer geometry
(layers_to_transform 0–79, r=64) and is not interchangeable with other JUWEL
bases.
python
from peft import PeftModelmodel = PeftModel.from_pretrained(model, "VextLabsinc/gem-ruby")
Status: PENDING. No benchmark numbers are published for this model yet, and that is a real absence rather than a hidden result.
Until 2026-07-28 the GEM adapters named a 144-layer base in their cards while being native to 80 layers, so nothing in the release was loadable as trained and no valid benchmark could be produced. Publishing Beryl is what makes the fleet runnable. Structural compatibility is verified — all 560 target modules per adapter resolve against this model's tensor index — but output-quality evaluation requires GPU time and has not been completed.
When numbers arrive they will be BF16, measured on our own weights on our own hardware, with per-sample audit trails. We do not publish a score we cannot reproduce, and we will not present quantized screening runs as headline results.
Research, evaluation, and authorized professional workflows where a human remains responsible for outcomes, and where the deployer supplies their own safety layer.
See USE_POLICY.md. In short: no unauthorized system access, no malware or fraud,
no CSAM, no weapons development, nothing violating applicable law. Publishing weights
is not permission to break the law, and Vext Labs does not endorse misuse.
Provided AS IS under LICENSE. To the maximum extent permitted by law,
Vext Labs, Inc. disclaims all warranties and is not liable for damages arising from
use or misuse. You are solely responsible for lawful, authorized use.