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

NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX

Original NOESIS MLX DubCompress specialist for multilingual dialogue compression, translation post-processing, dubbing transformations, and fitting translated text into fixed subtitle or speech-duration slots. This repository contains the MLX safetensors conversion of the BF16-trained NOESIS-Qwopus3.5-9B-DubCompress-v3.5 model for Apple Silicon inference through mlx-lm.

This is the MLX counterpart of the published NOESIS-Qwopus3.5-9B-DubCompress-v3.5-BF16 model card. The BF16 repository is the primary Transformers/GGUF distribution; this repository is the local MLX distribution of the same original NOESIS-trained specialist.

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

Property	Value
Model role	Dubbing compression, translation post-processing, and subtitle-slot fitting
Architecture	Qwen3_5ForCausalLM, qwen3_5
Parameters	Approximately 9B
Layers / hidden size	32 / 4096
Intermediate size	12288
Context limit in config	262,144 positions
Stored dtype	BF16 configuration; MLX safetensors
Weight shards	4 files, approximately 17.93 GB total
Tokenizer	Qwen3.5 tokenizer with chat_template.jinja
Runtime	mlx-lm on Apple Silicon
NOESIS version	v16.1
Release date	2026-08-26

Language support

The Qwen3.5 family advertises expanded coverage of 201 languages and dialects. The upstream Qwen3 language list below is the transparent documented baseline used for this derivative; the published Qwen3.5 count is broader than the enumerated list.

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.

Language coverage is inherited from the Qwen3.5/Qwen3 family and should not be interpreted as identical compression quality for every language. NOESIS production use must validate semantic preservation, target-language naturalness, subtitle length, syllable density, and timing budget.

Repository contents

NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX/
├── README.md
├── model-00001-of-00004.safetensors  # MLX shard 1
├── model-00002-of-00004.safetensors  # MLX shard 2
├── model-00003-of-00004.safetensors  # MLX shard 3
├── model-00004-of-00004.safetensors  # MLX shard 4
├── model.safetensors.index.json
├── config.json
├── generation_config.json
├── tokenizer.json
├── tokenizer_config.json
└── chat_template.jinja

The four shards and tokenizer files must remain together. Do not rename shards or remove model.safetensors.index.json; the MLX loader uses the index to reconstruct the complete model.

How to run with MLX-LM

Important platform note

mlx-lm is designed for Apple Silicon and uses the MLX backend. The B:\Downloads\... path is the Windows storage location of this repository; it is not a native Windows execution path for MLX. Copy or sync the repository to an Apple Silicon Mac, then use a macOS path such as /Users/ilia/Models/NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX.

Install

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip mlx-lm

Single prompt generation

DubCompress should receive the source or draft translation together with explicit constraints. Do not ask it to compress text without telling it the target language, the available duration or character budget, and the meaning that must be preserved.

python -m mlx_lm.generate \
  --model /Users/ilia/Models/NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX \
  --prompt "Target language: Russian. Slot budget: 4.2 seconds. Preserve names and meaning. Compress this dialogue to fit: Welcome to the Sanctuary, Miss Amelia." \
  --max-tokens 256 \
  --temp 0.3

For deterministic pipeline trials, use a low temperature and a bounded token limit. For long subtitle batches, reduce --prefill-step-size if memory spikes and set --max-kv-size to a deliberate bound.

Chat mode

python -m mlx_lm.chat \
  --model /Users/ilia/Models/NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX

Use the supplied chat_template.jinja through mlx-lm. A production prompt should identify the source text, target language, duration/character budget, forbidden changes, and the expected output format.

OpenAI-compatible local server

python -m mlx_lm.server \
  --model /Users/ilia/Models/NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX \
  --host 127.0.0.1 \
  --port 8080

Request it from another terminal:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "system", "content": "You are NOESIS DubCompress. Preserve meaning, names, tone, and target-language naturalness while fitting the stated slot."},
      {"role": "user", "content": "Target language: Spanish. Slot: 3.5 seconds. Return only the compressed line. Source: Welcome to the Sanctuary, Miss Amelia."}
    ],
    "temperature": 0.3,
    "max_tokens": 256
  }'

The MLX-LM server provides a basic OpenAI-compatible local endpoint. It is intended for local development and pipeline experiments and should not be exposed directly to the public Internet.

Python API

from mlx_lm import load, generate

model_dir = "/Users/ilia/Models/NOESIS-Qwopus3.5-9B-DubCompress-v3.5-MLX"
model, tokenizer = load(model_dir)

messages = [
    {
        "role": "system",
        "content": (
            "You are NOESIS DubCompress. Preserve meaning, names, tone, "
            "and target-language naturalness while fitting explicit timing constraints."
        ),
    },
    {
        "role": "user",
        "content": (
            "Target language: Russian. Slot budget: 4.2 seconds. "
            "Return only the compressed dialogue line. "
            "Source: Welcome to the Sanctuary, Miss Amelia."
        ),
    },
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
answer = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=256,
    temp=0.3,
    verbose=True,
)
print(answer)

NOESIS dubbing workflow

DubCompress is not a generic summarizer. It is a constrained language transformation stage in the NOESIS dubbing pipeline:

source transcript / translation
  → target-language instruction
  → slot duration, syllable, or character budget
  → DubCompress candidate
  → semantic and named-entity check
  → multilingual ASR / pronunciation check
  → timing and supervisor gates
  → TTS and sample-accurate assembly

The model must never be allowed to silently delete names, negation, safety-critical facts, or plot information merely to meet a duration budget. If the budget is impossible, return a warning or escalate to the configured subtitle fallback rather than accepting a semantically damaged line.

Task	Recommended model
Fast sanity check and compact verdict	Inspector 0.8B MLX
Translation compression and subtitle-slot fitting	DubCompress 9B MLX
Long-context review and final response judgment	Supervisor 9B MLX
Voice generation	OmniVoice / Talker route, not DubCompress

Runtime and memory notes

Item	Guidance
Weights	Approximately 17.93 GB across four safetensors shards.
Unified memory	Leave additional memory for the OS, tokenizer, and KV cache; required memory is higher than the raw file size.
Context	262,144 positions in configuration; practical context depends on Mac unified memory and KV-cache settings.
Sampling	Use bounded max_tokens; low temperature is recommended for constrained rewriting.
Chat format	Preserve and use chat_template.jinja.
Server	Bind to 127.0.0.1 for local-only use.
Quality	Validate semantic preservation, target-language quality, and fit after every model call.

NOESIS sealed rules

• R-DUB-COMPRESS-CONSTRAINED: every compression request must provide a target language and explicit slot or length constraint.
• R-DUB-COMPRESS-SEMANTIC-PRESERVE: names, negation, plot facts, and speaker intent must survive compression.
• R-DUB-COMPRESS-ASR-GATE: compressed candidates must pass multilingual ASR and pronunciation checks before TTS.
• R-DUB-COMPRESS-SUPERVISOR-GATE: use Supervisor for difficult or ambiguous candidates; do not accept low-confidence rewrites silently.
• R-MLX-LOCALHOST: the MLX-LM server is a local development service and must not be exposed without a separate secure gateway.
• R-CHAT-TEMPLATE-PRESERVE: preserve the tokenizer and shipped chat template with all weight shards.
• R-APACHE-CLEAN: preserve the applicable upstream and NOESIS license notices.

References

License

See the repository license and upstream notices. This MLX model is an original NOESIS-trained artifact released by AMAImedia; redistribution must preserve the model metadata and attribution.