AMAImedia

Qwen3-VL-2B-UI-Venus-NOESIS-BF16

README

License: apache-2.0

Summary

End-to-end GUI agent for autonomous UI navigation and element grounding. Trained via 4-stage pipeline (Mid-Train → Offline-RL → Online-RL → Model-Merge) on 10B GUI tokens across 30+ datasets. SOTA on ScreenSpot-Pro 57.7%, VenusBench-GD, OSWorld-G, AndroidWorld/Lab, WebVoyager.

Use case inside NOESIS

GUI agent / screenshot understanding / UI element grounding. Inside NOESIS this is an out-of-scope sibling — kept for general agent research, not the dubbing pipeline.

What changed vs upstream

Table with columns: Aspect, Upstream, This bundle
Aspect	Upstream	This bundle
Floating-point storage dtype	FP16	bfloat16
`config.json` `torch_dtype`	as-is	`bfloat16`
`model.safetensors.index.json` `total_size`	as-is	recomputed
Tokenizer / chat template / modeling code	as-is	unchanged
Number of parameters	as-is	unchanged
Value-level transformation beyond dtype cast	—	none
Disk size	4.6 GB	4.6 GB

Architecture

Table with columns: Property, Value
Property	Value
Immediate parent	`inclusionAI/UI-Venus-1.5-2B`
Architecture	`Qwen3VLForConditionalGeneration`
Architecture base / lineage	Qwen3-VL-2B (4-stage post-training: Mid-Train 10B tok GUI → Offline-RL → Online-RL → Model-Merge)
Parameters	~2B (dense)
Hidden size	see text_config in config.json
Num hidden layers	see text_config in config.json
Attention heads / KV heads

Repack tooling

CPU-only sharded repack via repack_fp32_to_bf16.py — reads each shard with safetensors.safe_open, casts floating-point tensors to torch.bfloat16, rewrites the shard, updates the index manifest. No GPU involvement, no value-level transformation beyond the IEEE-754 dtype cast.

Performance reference (RTX 3060 laptop, NVMe SSD):

Single 5 GB FP32 shard cast → ~28-40 sec
Full 4.6 GB → 4.6 GB in 1 pass, sharded

Use cases (for the BF16 bundle)

✅ LoRA / DoRA / IA³ fine-tuning that requires a dtype=torch.bfloat16 base
✅ Bitsandbytes NF4 / AWQ-INT4 / GPTQ quantization (these tools prefer BF16 input)
✅ Inference on Ampere+ / MI200+ hardware with native BF16 support
✅ KD-teacher (forward-only) where BF16 storage saves bandwidth
❌ Full-parameter fine-tuning of weights — use FP32/BF16 master weights pattern; storage dtype alone is insufficient

Quick start

python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "AMAImedia/UI-Venus-1.5-2B-NOESIS-BF16"

tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
).eval()

Sealed rules (NOESIS DHCF-FNO)

R-DTYPE-REPACK-BF16 — pure IEEE-754 dtype cast from FP16 to bfloat16. No value-level transformation, no LoRA merge, no architectural change. Equivalent to loading upstream with dtype=torch.bfloat16 and saving, but materialised on disk.
R-APACHE-CLEAN — upstream Apache License 2.0 preserved end-to-end via the LICENSE file in this repo. AMAImedia adds only a derivative-work notice for the repack step.
R-NO-VALUE-TRANSFORM — no fine-tuning, no distillation, no merge has been applied between upstream and this repo. Outputs are bit-for-bit equivalent up to the precision difference of the dtype cast.

License & attribution

This bundle inherits Apache License 2.0 from inclusionAI/UI-Venus-1.5-2B. Original model card, citation, and attribution from upstream apply without modification. See LICENSE in this repo for the complete text plus the NOESIS derivative-work NOTICE.

Citation

bibtex
@misc{noesis2026uivenus152bnoesisbf16bf16,
  title  = {NOESIS DHCF-FNO :: UI-Venus-1.5-2B-NOESIS-BF16 — BF16 dtype-repack derivative},
  author = {Bolotnikov, Ilia and AMAImedia},
  year   = {2026},
  note   = {BF16 dtype-repack of inclusionAI/UI-Venus-1.5-2B for LoRA / PEFT
            compatibility. 4.6 GB on disk, Apache License 2.0
            preserved end-to-end.},
  url    = {https://huggingface.co/AMAImedia/UI-Venus-1.5-2B-NOESIS-BF16}
}

Please also cite the upstream model when using this bundle. See the upstream README and LICENSE in this repo for citation requirements.

Author

Founder: Ilia Bolotnikov
Organization: AMAImedia.com
X (Twitter): @AMAImediacom
LinkedIn: Ilia Bolotnikov
Telegram: @AMAImediacom
NOESIS version: v15.8
Repack date: 2026-05-19
HF repo: AMAImedia/UI-Venus-1.5-2B-NOESIS-BF16

Produced 2026-05-19 by NOESIS DHCF-FNO v15.8 — AMAImedia.com

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

AMAImedia

Model Tree

Base

inclusionAI/UI-Venus-1.5-2B

Fine-tuned

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer