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: v15.9
Release date: 2026-05

Last updated: 2026-08-28

NOESIS-MiniCPM-V-4.6-SigLIP2-ONNX-INT8

INT8 ONNX deployment variant of the SigLIP2 vision encoder extracted from MiniCPM-V-4.6. This auxiliary vision feature extractor supports cinema-dubbing pipeline diagnostics such as poster-art preparation, scene-reference-frame tagging, and key-frame cross-checks. It is not wired into the audio dubbing core under R-DUBBING-FILM-SCOPE / SCOPE LOCK.

NOESIS provenance

Property	Value
Bundle	`NOESIS-MiniCPM-V-4.6-SigLIP2-ONNX-INT8`
Parent bundle	`NOESIS-3.5B-A0.5B-DUBBING-FILM`
BF16 source	`../NOESIS-MiniCPM-V-4.6-SigLIP2-BF16` (sibling)
Quantization	INT8 ONNX for low-VRAM / CPU deployment via onnxruntime
NOESIS role	Auxiliary SigLIP2 image-feature extractor
NOESIS version	v15.9
Original release	2026-05
Last updated	2026-08-28

================================================================================ Architecture

Property	Value
Component	SigLIP2 vision encoder (from MiniCPM-V-4.6)
Precision	INT8 ONNX (per-channel)
File	`vision_encoder_int8.onnx` (~465 MB, no external data)
Source bundle	../NOESIS-MiniCPM-V-4.6-SigLIP2-BF16 (~996 MB BF16)
Compression vs BF16	~2.14×
Runtime	onnxruntime (CPU / CUDA / DirectML)
Output	image feature embeddings
License	Apache-2.0

================================================================================ Bundle Contents

.
├── README.md
├── LICENSE                          # Apache-2.0 + NOESIS notice
├── config.json                      # vision encoder config
├── preprocessor_config.json         # image preprocessor (resize/normalize)
├── chat_template.jinja              # carried over from MiniCPM-V chat template
└── vision_encoder_int8.onnx         # INT8 quantized weights (~465 MB)

================================================================================ Family

Bundle	Format	Size	Use
NOESIS-MiniCPM-V-4.6-SigLIP2-BF16	BF16 safetensors	~996 MB	source-of-truth
NOESIS-MiniCPM-V-4.6-SigLIP2-ONNX-INT8 ← this bundle	INT8 ONNX	~465 MB	deployment

================================================================================ Quick Start (onnxruntime)

import onnxruntime as ort
from transformers import AutoImageProcessor
from PIL import Image
import numpy as np

BUNDLE = ("B:/Downloads/Portable/NOESIS-VC-ONE/models/llm/"
          "NOESIS-3.5B-A0.5B-DUBBING-FILM/"
          "NOESIS-MiniCPM-V-4.6-SigLIP2-ONNX-INT8")

proc = AutoImageProcessor.from_pretrained(BUNDLE)
sess = ort.InferenceSession(f"{BUNDLE}/vision_encoder_int8.onnx",
                            providers=["CPUExecutionProvider"])

img = Image.open("scene.png").convert("RGB")
pixel_values = proc(img, return_tensors="np")["pixel_values"]
features = sess.run(None, {"pixel_values": pixel_values})[0]

================================================================================ NOESIS Sealed Rules

R-APACHE-CLEAN Upstream MiniCPM-V Apache-2.0 preserved verbatim in LICENSE. Commercial use permitted.

R-VENDORED-INTERNAL Internal vendor copy inside parent NOESIS-3.5B-A0.5B-DUBBING-FILM bundle.

R-DUBBING-FILM-SCOPE Auxiliary vision encoder — NOT in the core audio dubbing pipeline (SCOPE LOCK 2026-05-15: NOESIS core scope = audio dubbing).

R-NOESIS-FINAL-ARTIFACT-PATHS Canonical path: models/llm/NOESIS-3.5B-A0.5B-DUBBING-FILM/NOESIS-MiniCPM-V-4.6-SigLIP2-ONNX-INT8/

R-NEVER-DELETE-WITHOUT-EXPLICIT-CONSENT Sealed 2026-05-21.

================================================================================ Upstream Citation

Upstream: https://huggingface.co/openbmb/MiniCPM-V-4_5 GitHub: https://github.com/OpenBMB/MiniCPM-V

================================================================================