NOESIS / AMAImedia

Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

Founder: Ilia Bolotnikov
Organization: AMAImedia.com
X (Twitter): @AMAImediacom
LinkedIn: Ilia Bolotnikov
Telegram: @djbionicl
NOESIS version: v16.1
Release date: 2026-08-26

Language support

This Qwen3-derived model follows the official Qwen3 language coverage (119 languages and dialects):

English, French, Portuguese, German, Romanian, Swedish, Danish, Bulgarian, Russian, Czech, Greek, Ukrainian, Spanish, Dutch, Slovak, Croatian, Polish, Lithuanian, Norwegian Bokmål, Norwegian Nynorsk, Persian, Slovenian, Gujarati, Latvian, Italian, Occitan, Nepali, Marathi, Belarusian, Serbian, Luxembourgish, Venetian, Assamese, Welsh, Silesian, Asturian, Chhattisgarhi, Awadhi, Maithili, Bhojpuri, Sindhi, Irish, Faroese, Hindi, Punjabi, Bengali, Oriya, Tajik, Eastern Yiddish, Lombard, Ligurian, Sicilian, Friulian, Sardinian, Galician, Catalan, Icelandic, Tosk Albanian, Limburgish, Dari, Afrikaans, Macedonian, Sinhala, Urdu, Magahi, Bosnian, Armenian; Chinese (Simplified Chinese, Traditional Chinese, Cantonese), Burmese; Arabic (Standard, Najdi, Levantine, Egyptian, Moroccan, Mesopotamian, Ta’izzi-Adeni, Tunisian), Hebrew, Maltese; Indonesian, Malay, Tagalog, Cebuano, Javanese, Sundanese, Minangkabau, Balinese, Banjar, Pangasinan, Iloko, Waray (Philippines); Tamil, Telugu, Kannada, Malayalam; Turkish, North Azerbaijani, Northern Uzbek, Kazakh, Bashkir, Tatar; Thai, Lao; Finnish, Estonian, Hungarian; Vietnamese, Khmer; Japanese, Korean, Georgian, Basque, Haitian, Papiamento, Kabuverdianu, Tok Pisin, Swahili.

--- library_name: transformers license: apache-2.0 license_link: LICENSE.md base_model: - Tongyi/MAI-UI-2B pipeline_tag: image-text-to-text tags: - nf4 - bnb-4bit - bitsandbytes - quantization - apache-2.0 - commercial-ok - noesis - noesis-ui-agent - noesis-qwen3-vl - mai-ui - gui - agent - multimodal - browser-dom-automation - qwen3-vl - dhcf-fno - amaimedia - vendored-internal - qwen3 - qwen3-119-languages - supports-119-languages - multilingual - language-support language: - en - zh

NOESIS-Qwen3-VL-2B-MAI-UI-NF4 (NOESIS DHCF-FNO bundle)

NF4 quantization derivative of Tongyi/MAI-UI-2B — Real-World Centric Foundation GUI Agent (2B variant of the MAI-UI family: 2B / 8B / 32B / 235B-A22B). NF4-quantized via bitsandbytes 0.49.2 (double_quant + bf16 compute) from the intermediate Tongyi-MAI-UI-2B-BF16 AMAImedia BF16 repack.

Used inside the NOESIS DHCF-FNO stack as the SECONDARY 2B agent on the public ui-agent.amaimedia.com subdomain (browser DOM automation). The PRIMARY agent is the 8B NF4 sibling (NOESIS-Qwen3-VL-8B-MAI-UI-NF4) per R-AGENT-PRIMARY-MAI-UI-8B-NF4.

✅ APACHE 2.0 — COMMERCIAL USE PERMITTED. End-to-end clean lineage (Alibaba Cloud / Qwen Team Apache 2.0 → Alibaba Tongyi MAI-UI Apache 2.0 → AMAImedia BF16 repack Apache 2.0 → AMAImedia NF4 Apache 2.0). Standard transformers.from_pretrained loading with device_map={"": 0} (NF4 requirement per CLAUDE.md GOLDEN RULE 2).

Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

Quantization date: 2026-05-21 08:50:12

NOESIS role — secondary 2B agent on ui-agent.amaimedia.com

Browser DOM automation agent mounted on ui-agent.amaimedia.com (Phase 2 desktop agent / auto-clipper UI nav subdomain). This 2B variant is the lightweight fallback for the canonical 8B NF4 sibling and the staging ground for LoRA recipe validation before committing GPU hours on 8B.

ui-agent.amaimedia.com (browser DOM automation)
        │
        ├── PRIMARY  : NOESIS-Qwen3-VL-8B-MAI-UI-NF4  (~5 GB VRAM)
        │              R-AGENT-PRIMARY-MAI-UI-8B-NF4
        │
        └── SECONDARY: NOESIS-Qwen3-VL-2B-MAI-UI-NF4  (this, ~1.6 GB VRAM)
                       • low-VRAM fallback
                       • LoRA recipe staging
                       • parallel-environment scaling tests

Property	Value
Immediate parent	`Tongyi-MAI-UI-2B-BF16` (AMAImedia BF16 repack of `Tongyi/MAI-UI-2B`)
Upstream lineage	`Qwen/Qwen3-VL-2B` (Apache 2.0) → `Tongyi/MAI-UI-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	1.57 GB (single safetensors)
VRAM target (inference)	1.5 GB ✅ RTX 3060 6 GB
VRAM peak (load)	1.6 GB
License	Apache 2.0 (commercial-ok)

Upstream Tongyi MAI-UI documentation (verbatim)

MAI-UI: Real-World Centric Foundation GUI Agents.

📖 Background

The development of GUI agents could revolutionize the next generation of human-computer interaction. Motivated by this vision, we present MAI-UI, a family of foundation GUI agents spanning the full spectrum of sizes, including 2B, 8B, 32B, and 235B-A22B variants. We identify four key challenges to realistic deployment: the lack of native agent–user interaction, the limits of UI-only operation, the absence of a practical deployment architecture, and brittleness in dynamic environments. MAI-UI addresses these issues with a unified methodology: a self-evolving data pipeline that expands the navigation data to include user interaction and MCP tool calls, a native device–cloud collaboration system that routes execution by task state, and an online RL framework with advanced optimizations to scale parallel environments and context length.

🏆 Results

Grounding

MAI-UI establishes new state-of-the-art across GUI grounding and mobile navigation.

On grounding benchmarks, it reaches 73.5% on ScreenSpot-Pro, 91.3% on MMBench GUI L2, 70.9% on OSWorld-G, and 49.2% on UI-Vision, surpassing Gemini-3-Pro and Seed1.8 on ScreenSpot-Pro.

Mobile Navigation

On mobile GUI navigation, it sets a new SOTA of 76.7% on AndroidWorld, surpassing UI-Tars-2, Gemini-2.5-Pro and Seed1.8. On MobileWorld, MAI-UI obtains 41.7% success rate, significantly outperforming end-to-end GUI models and competitive with Gemini-3-Pro based agentic frameworks.

Online RL

Our online RL experiments show significant gains from scaling parallel environments from 32 to 512 (+5.2 points) and increasing environment step budget from 15 to 50 (+4.3 points).

Device-Cloud Collaboration

The device-cloud collaboration framework can dynamically select on-device or cloud execution based on task execution state and data sensitivity. It improves on-device performance by 33% and reduces cloud API calls by over 40%.

Quantization details (NOESIS-side)

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\Tongyi-MAI-UI-2B-BF16`
Output disk size	1.57 GB (single safetensors)
VRAM target (inference)	1.5 GB
VRAM peak (load)	1.6 GB
Quant date	2026-05-21 08:50:12

Standard from_pretrained path — no custom workarounds needed for Qwen3-VL family. Vision tower retained (required for screenshot grounding tasks).

Quick start

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

bundle = "B:/Downloads/Portable/NOESIS-VC-ONE/models/llm/NOESIS-Qwen3-VL-2B-MAI-UI-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 example
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "screenshot.png"},
            {"type": "text",  "text": "Click the 'Submit' 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 Submit button

NOESIS ui-agent.amaimedia.com wiring

Endpoint	Role	Backend
`ui-agent.amaimedia.com` (PRIMARY tier)	Browser DOM automation, full SOTA quality	`NOESIS-Qwen3-VL-8B-MAI-UI-NF4` (sibling, ~5 GB VRAM)
`ui-agent.amaimedia.com` (SECONDARY tier)	Low-VRAM fallback, LoRA staging, parallel-env scaling	THIS bundle (~1.6 GB VRAM)
`ui-agent.amaimedia.com` (FALLBACK tier)	Alternative training pipeline (4-stage RFT)	`NOESIS-Qwen3-VL-2B-UI-Venus-NF4` (sibling, ~3.5 GB VRAM peak)

Sealed rules (NOESIS DHCF-FNO)

R-APACHE-CLEAN — Apache 2.0 preserved end-to-end (Qwen Team → Alibaba Tongyi → AMAImedia BF16 repack → AMAImedia NF4 quant).
R-NF4-DEVICE-MAP-EXPLICIT — must load with device_map={"": 0}; never device_map="auto" with NF4 (CLAUDE.md GOLDEN RULE 2).
R-AGENT-PRIMARY-MAI-UI-8B-NF4 — the 8B NF4 sibling is the PRIMARY ui-agent.amaimedia.com agent; this 2B variant is SECONDARY (fallback + LoRA validation).
R-MAI-UI-SOTA-AGENT — SOTA on ScreenSpot-Pro 73.5%, MMBench GUI L2 91.3%, OSWorld-G 70.9%, AndroidWorld 76.7%, MobileWorld 41.7%.
R-DEVICE-CLOUD-COLLAB-CAPABLE — native device-cloud collaboration: +33% on-device, -40% cloud API calls per upstream report.
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 alongside LICENSE.md.
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 — can KD with sibling Qwen3 text models if needed.

NOESIS provenance

Step	Source / output
Base architecture	`Qwen/Qwen3-VL-2B` (© Alibaba Cloud / Qwen Team 2025-2026, Apache 2.0)
GUI agent fine-tune	`Tongyi/MAI-UI-2B` (© Alibaba Tongyi 2026, Apache 2.0)
BF16 dtype-repack (intermediate)	`Tongyi-MAI-UI-2B-BF16` (© AMAImedia 2026, Apache 2.0)
NF4 quantization	`bitsandbytes` 0.49.2 + double-quant + bf16 compute
Local file	`model.safetensors` (1.57 GB) + `config.json` + processor + tokenizer
Quant date	2026-05-21 08:50:12
NOESIS version	v15.8
Production endpoint	`ui-agent.amaimedia.com` (Phase 2 subdomain)

Reference docs:

NOESIS CLAUDE.md GOLDEN RULE 2 (NF4 device_map={"":0})
NOESIS sealed rule R-AGENT-PRIMARY-MAI-UI-8B-NF4
NOESIS_NF4_MANIFEST.json in this folder

License

Apache License 2.0. Qwen3-VL base architecture © Alibaba Cloud / Qwen Team. Tongyi MAI-UI 2B fine-tune © Alibaba Tongyi. 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 and LICENSE.md in this folder for the full Apache 2.0 text plus the NOESIS attribution / NOTICE block.

Quantization date: 2026-05-21 08:50:12
Parent BF16 source: Tongyi-MAI-UI-2B-BF16 (D:\models\vlm-gui-mot)
Vendored component: NOESIS-Qwen3-VL-2B-MAI-UI-NF4 (Apache 2.0)

Produced 2026-08-26 by NOESIS DHCF-FNO v15.8 — AMAImedia.com

Downloads last month: 110

Safetensors

Model size

2B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including AMAImedia/Qwen3-VL-2B-MAI-UI-NOESIS-NF4

NOESIS-RePack-Quant-Models

Collection

27 items • Updated about 3 hours ago