GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
AMAImedia
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
AMAImedia
Model Tree
inclusionAI/UI-Venus-1.5-2B
Input Modalities
Output Modalities
Supported Functionality
Browser DOM automation agent mounted on ui-agent.amaimedia.com
(Phase 2 desktop agent / auto-clipper UI nav subdomain) as the
FALLBACK tier in a 3-tier hierarchy. Provides cross-validation
against MAI-UI's grounding output through a fundamentally different
training pipeline (4-stage RFT vs self-evolving data + device-cloud
collab).
text
ui-agent.amaimedia.com (browser DOM automation) │ ├── PRIMARY : NOESIS-Qwen3-VL-8B-MAI-UI-NF4 (Tongyi MAI-UI 8B, ~5 GB VRAM) │ R-AGENT-PRIMARY-MAI-UI-8B-NF4 │ ├── SECONDARY: NOESIS-Qwen3-VL-2B-MAI-UI-NF4 (Tongyi MAI-UI 2B, ~1.6 GB VRAM) │ low-VRAM fallback for primary │ └── FALLBACK : NOESIS-Qwen3-VL-2B-UI-Venus-NF4 (this, Inclusion AI Venus 1.5 2B) • alternative 4-stage RFT pipeline • cross-validation against MAI-UI • ~1.2 GB target / 3.45 GB load peak VRAMThe Venus variant exists as the cross-validation track — when MAI-UI gives ambiguous click coordinates or fails to ground an element, Venus provides a second opinion from an entirely separate training pipeline (RFT-based, not RL-from-environment).
| Property | Value |
|---|---|
| Immediate parent | UI-Venus-1.5-2B-NOESIS-BF16 (AMAImedia BF16 repack of inclusionAI/UI-Venus-1.5-2B) |
| Upstream lineage | Qwen/Qwen3-VL-2B (Apache 2.0) → inclusionAI/UI-Venus-1.5-2B (Apache 2.0) → AMAImedia BF16 repack → AMAImedia NF4 |
| Architecture | Qwen3VLForConditionalGeneration (multimodal, vision tower retained) |
| Text hidden | 2 048 / 28 layers / 16 heads (GQA 2 : 1, 8 kv heads) |
| Vision tower | depth 24, hidden 1024, patch 16, deepstack at layers [5,11,17] |
| Vocab size | 151 936 |
| Context | 262 144 (mRoPE [24,20,20] interleaved, rope_theta 5M) |
| Format | NF4 (bnb 4-bit, double-quant, bf16 compute) |
| Bundle size on disk | 2.19 GB (single safetensors) |
| VRAM target (inference) | 1.2 GB ✅ RTX 3060 6 GB |
| VRAM peak (load) | 3.45 GB |
| License | Apache 2.0 (commercial-ok) |
| Project page | https://ui-venus.github.io/UI-Venus-1.5 |
| Papers | arxiv:2602.09082 (Venus 1.5), arxiv:2508.10833 (Venus RFT) |
UI-Venus-1.5 is a unified end-to-end GUI Agent designed for robust real-world applications. The model family includes two dense variants (2B / 8B) and one MoE variant (30B-A3B). This folder hosts the 2B dense variant, NF4-quantized.
text
Stage 1 — Mid-Training 10B tokens across 30+ GUI datasets Foundational GUI semantics Stage 2 — Offline-RL Task-specific optimization: • grounding (ScreenSpot / OSWorld-G / VenusBench-GD) • mobile (AndroidWorld / AndroidLab / VenusBench-Mobile) • web (WebVoyager / OSWorld-W) Stage 3 — Online-RL Full-trajectory rollouts for long-horizon dynamic navigation RFT (Reinforcement Fine-Tuning) per arXiv 2508.10833 Stage 4 — Model Merge Unifying specialists into single deployable checkpoint| Benchmark | 30B-A3B variant | This 2B variant (proportional) |
|---|---|---|
| ScreenSpot-Pro | 69.6% | 57.7% |
| VenusBench-GD | 75.0% | (scales with size) |
| OSWorld-G-R | 76.4% | (scales with size) |
| OSWorld-G | 70.6% | (scales with size) |
| UI-Vision | 54.7% | (scales with size) |
| AndroidWorld | 77.6% | (scales with size) |
| AndroidLab | 55.1% / 68.1% | (scales with size) |
| VenusBench-Mobile | 21.5% | (scales with size) |
| WebVoyager | 76.0% | (scales with size) |
Numbers per upstream Venus Team report; 2B variant proportionally lower per "Consistent Scaling Gains" note in upstream README.
| Parameter | Value |
|---|---|
| Library | bitsandbytes 0.49.2 |
| Method | NF4 (Normalized Float 4-bit) |
bnb_4bit_use_double_quant | True (saves ~5% via nested quant) |
bnb_4bit_compute_dtype | bfloat16 |
| Device map | {"": 0} (R-NF4-DEVICE-MAP-EXPLICIT) |
| Source dir | D:\models\vlm-gui-mot\UI-Venus-1.5-2B-NOESIS-BF16 |
| Output disk size | 2.19 GB (single safetensors) |
| VRAM target (inference) | 1.2 GB |
| VRAM peak (load) | 3.45 GB |
| Quant date | 2026-05-21 08:32:59 |
Higher load peak (3.45 GB vs MAI-UI 2B 1.6 GB) reflects Venus's internal layer-precision retention strategy — the working set settles to 1.2 GB after warmup, but initialization touches more parameters in higher precision.
python
import torchfrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration bundle = "B:/Downloads/Portable/NOESIS-VC-ONE/models/llm/NOESIS-Qwen3-VL-2B-UI-Venus-NF4" processor = AutoProcessor.from_pretrained(bundle)model = Qwen3VLForConditionalGeneration.from_pretrained( bundle, device_map={"": 0}, # NEVER "auto" with NF4 torch_dtype=torch.bfloat16,).eval() # Browser DOM screenshot grounding examplemessages = [ { "role": "user", "content": [ {"type": "image", "image": "screenshot.png"}, {"type": "text", "text": "Click the 'Subscribe' button."}, ], },]inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_tensors="pt",).to(0)with torch.no_grad(): outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(processor.decode(outputs[0], skip_special_tokens=True))# → predicted bounding box / click coordinates for the Subscribe button| Endpoint tier | Backend | VRAM | Role |
|---|---|---|---|
| PRIMARY | NOESIS-Qwen3-VL-8B-MAI-UI-NF4 | ~5 GB | Canonical agent — full SOTA quality |
| SECONDARY | NOESIS-Qwen3-VL-2B-MAI-UI-NF4 | ~1.6 GB | Low-VRAM fallback for low-spec clients |
| FALLBACK | THIS bundle | ~1.2 GB target / 3.45 GB load peak | Cross-validation via alternative 4-stage RFT pipeline |
When to invoke FALLBACK (Venus) over PRIMARY/SECONDARY (MAI-UI):
R-APACHE-CLEAN — Apache 2.0 preserved end-to-end (Qwen Team → Inclusion AI Venus Team → AMAImedia BF16 repack → AMAImedia NF4 quant).R-NF4-DEVICE-MAP-EXPLICIT — must load with device_map={"": 0}; never device_map="auto" with NF4.R-AGENT-PRIMARY-MAI-UI-8B-NF4 — MAI-UI 8B NF4 is PRIMARY, MAI-UI 2B NF4 is SECONDARY, this Venus 2B is FALLBACK (alternative pipeline cross-validation).R-UI-VENUS-FALLBACK-TO-MAI-UI — Venus is fallback / cross-validation track; MAI-UI is canonical agent path.R-VENUS-15-4STAGE-PIPELINE — 4-stage post-training: Mid-Train (10B GUI tok, 30+ datasets) → Offline-RL → Online-RL → Model Merge.R-VENUS-RFT-TRAINED — Reinforcement Fine-Tuning (RFT) per arXiv 2508.10833.R-QWEN3-VL-MROPE-INTERLEAVED — mRoPE [24, 20, 20] interleaved with rope_theta 5M (text); 256K context capable.R-UI-AGENT-PRODUCT-SCOPE — mounted on ui-agent.amaimedia.com (browser DOM automation), NOT the dubbing pipeline core path.R-VENDORED-INTERNAL — plain LICENSE preserved (BF16-tier NOTICE blocks) alongside LICENSE.md (NF4-tier NOTICE).R-THIRD-PARTY-WRAPPERS-ONLY — Phase 1 SCOPE LOCK — third-party + wrappers only, no own training.R-VISION-TOWER-RETAINED — full Qwen3-VL ViT preserved (depth 24, deepstack at [5,11,17]) — required for screenshot grounding.R-QWEN-VOCAB-151936 — compatible within Qwen3 family.R-UI-AGENT-OUT-OF-SCOPE — NOT in NOESIS dubbing-pipeline core path. Reserved for Phase 2 desktop agent / auto-clipper UI navigation experiments.| Step | Source / output |
|---|---|
| Base architecture | Qwen/Qwen3-VL-2B (© Alibaba Cloud / Qwen Team 2025-2026, Apache 2.0) |
| GUI agent fine-tune | inclusionAI/UI-Venus-1.5-2B (© Inclusion AI / Ant Group Venus Team 2025-2026, Apache 2.0) |
| Training pipeline | 4-stage: Mid-Train (10B GUI tok) → Offline-RL → Online-RL → Merge (RFT per arXiv 2508.10833) |
| BF16 dtype-repack (intermediate) | UI-Venus-1.5-2B-NOESIS-BF16 (© AMAImedia 2026, Apache 2.0, 4.6 GB) |
| NF4 quantization | bitsandbytes 0.49.2 + double-quant + bf16 compute |
| Local file | model.safetensors (2.19 GB) + config.json + processor + tokenizer |
| Quant date | 2026-05-21 08:32:59 |
| NOESIS version | v15.8 |
| Renamed from | NOESIS-UI-Venus-2B-NF4 (per R-NOESIS-FOLDER-NAMING-PREFIX-QWEN3-VL) |
| Production endpoint | ui-agent.amaimedia.com (Phase 2 subdomain) |
Reference docs:
R-AGENT-PRIMARY-MAI-UI-8B-NF4NOESIS_NF4_MANIFEST.json in this folderbibtex
@misc{venusteam2026uivenus15technicalreport, title={UI-Venus-1.5 Technical Report}, author={Venus Team and Changlong Gao and Zhangxuan Gu and Yulin Liu and Xinyu Qiu and Shuheng Shen and Yue Wen and Tianyu Xia and Zhenyu Xu and Zhengwen Zeng and Beitong Zhou and Xingran Zhou and Weizhi Chen and Sunhao Dai and Jingya Dou and Yichen Gong and Yuan Guo and Zhenlin Guo and Feng Li and Qian Li and Jinzhen Lin and Yuqi Zhou and Linchao Zhu and Liang Chen and Zhenyu Guo and Changhua Meng and Weiqiang Wang}, year={2026}, eprint={2602.09082}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2602.09082}} @misc{gu2025uivenustechnicalreportbuilding, title={UI-Venus Technical Report: Building High-performance UI Agents with RFT}, author={Zhangxuan Gu and Zhengwen Zeng and Zhenyu Xu and Xingran Zhou and Shuheng Shen and Yunfei Liu and Beitong Zhou and Changhua Meng and Tianyu Xia and Weizhi Chen and Yue Wen and Jingya Dou and Fei Tang and Jinzhen Lin and Yulin Liu and Zhenlin Guo and Yichen Gong and Heng Jia and Changlong Gao and Yuan Guo and Yong Deng and Zhenyu Guo and Liang Chen and Weiqiang Wang}, year={2025}, eprint={2508.10833}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2508.10833}} @misc{noesis2026qwen3vl2buivenusnf4, title = {NOESIS DHCF-FNO :: Qwen3-VL-2B-UI-Venus NF4 — FALLBACK agent on ui-agent.amaimedia.com}, author = {Bolotnikov, Ilia and AMAImedia}, year = {2026}, note = {NF4 (bitsandbytes) quantization derivative of inclusionAI/UI-Venus-1.5-2B (via AMAImedia BF16 repack), Apache 2.0. 2.19 GB on disk, 1.2 GB VRAM target on RTX 3060.}, url = {https://amaimedia.com}}Apache License 2.0. Qwen3-VL base architecture © Alibaba Cloud / Qwen Team. UI-Venus-1.5-2B 4-stage RFT fine-tune © Inclusion AI / Ant Group (Venus Team). BF16 dtype-repack + NF4 quantization + NOESIS bundling + sealed-rule wiring: © AMAImedia (NOESIS DHCF-FNO project) 2026.
Commercial use is permitted subject to the standard Apache 2.0
preservation requirements (copyright + LICENSE + NOTICE-equivalent
attribution must travel with redistributions). See LICENSE (plain
text, BF16-tier NOTICE blocks) and LICENSE.md (Markdown, NF4-tier
NOTICE) in this folder for the full attribution chain.
UI-Venus-1.5-2B-NOESIS-BF16 (D:\models\vlm-gui-mot)inclusionAI/UI-Venus-1.5-2BProduced 2026-05-21 by NOESIS DHCF-FNO v15.8 — AMAImedia.com
ui-agent.amaimedia.com (browser DOM automation) │ ├── PRIMARY : NOESIS-Qwen3-VL-8B-MAI-UI-NF4 (Tongyi MAI-UI 8B, ~5 GB VRAM) │ R-AGENT-PRIMARY-MAI-UI-8B-NF4 │ ├── SECONDARY: NOESIS-Qwen3-VL-2B-MAI-UI-NF4 (Tongyi MAI-UI 2B, ~1.6 GB VRAM) │ low-VRAM fallback for primary │ └── FALLBACK : NOESIS-Qwen3-VL-2B-UI-Venus-NF4 (this, Inclusion AI Venus 1.5 2B) • alternative 4-stage RFT pipeline • cross-validation against MAI-UI • ~1.2 GB target / 3.45 GB load peak VRAMStage 1 — Mid-Training 10B tokens across 30+ GUI datasets Foundational GUI semanticsStage 2 — Offline-RL Task-specific optimization: • grounding (ScreenSpot / OSWorld-G / VenusBench-GD) • mobile (AndroidWorld / AndroidLab / VenusBench-Mobile) • web (WebVoyager / OSWorld-W)Stage 3 — Online-RL Full-trajectory rollouts for long-horizon dynamic navigation RFT (Reinforcement Fine-Tuning) per arXiv 2508.10833Stage 4 — Model Merge Unifying specialists into single deployable checkpointimport torchfrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration bundle = "B:/Downloads/Portable/NOESIS-VC-ONE/models/llm/NOESIS-Qwen3-VL-2B-UI-Venus-NF4" processor = AutoProcessor.from_pretrained(bundle)model = Qwen3VLForConditionalGeneration.from_pretrained( bundle, device_map={"": 0}, # NEVER "auto" with NF4 torch_dtype=torch.bfloat16,).eval() # Browser DOM screenshot grounding examplemessages = [ { "role": "user", "content": [ {"type": "image", "image": "screenshot.png"}, {"type": "text", "text": "Click the 'Subscribe' button."}, ], },]inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_tensors="pt",).to(0)with torch.no_grad(): outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(processor.decode(outputs[0], skip_special_tokens=True))# → predicted bounding box / click coordinates for the Subscribe button@misc{venusteam2026uivenus15technicalreport, title={UI-Venus-1.5 Technical Report}, author={Venus Team and Changlong Gao and Zhangxuan Gu and Yulin Liu and Xinyu Qiu and Shuheng Shen and Yue Wen and Tianyu Xia and Zhenyu Xu and Zhengwen Zeng and Beitong Zhou and Xingran Zhou and Weizhi Chen and Sunhao Dai and Jingya Dou and Yichen Gong and Yuan Guo and Zhenlin Guo and Feng Li and Qian Li and Jinzhen Lin and Yuqi Zhou and Linchao Zhu and Liang Chen and Zhenyu Guo and Changhua Meng and Weiqiang Wang}, year={2026}, eprint={2602.09082}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2602.09082}} @misc{gu2025uivenustechnicalreportbuilding, title={UI-Venus Technical Report: Building High-performance UI Agents with RFT}, author={Zhangxuan Gu and Zhengwen Zeng and Zhenyu Xu and Xingran Zhou and Shuheng Shen and Yunfei Liu and Beitong Zhou and Changhua Meng and Tianyu Xia and Weizhi Chen and Yue Wen and Jingya Dou and Fei Tang and Jinzhen Lin and Yulin Liu and Zhenlin Guo and Yichen Gong and Heng Jia and Changlong Gao and Yuan Guo and Yong Deng and Zhenyu Guo and Liang Chen and Weiqiang Wang}, year={2025}, eprint={2508.10833}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2508.10833}} @misc{noesis2026qwen3vl2buivenusnf4, title = {NOESIS DHCF-FNO :: Qwen3-VL-2B-UI-Venus NF4 — FALLBACK agent on ui-agent.amaimedia.com}, author = {Bolotnikov, Ilia and AMAImedia}, year = {2026}, note = {NF4 (bitsandbytes) quantization derivative of inclusionAI/UI-Venus-1.5-2B (via AMAImedia BF16 repack), Apache 2.0. 2.19 GB on disk, 1.2 GB VRAM target on RTX 3060.}, url = {https://amaimedia.com}}