Quality-tier dictation touch-up model for the Yooz suite. Qwen3.5 4B (6-bit g64 MLX, fused from the full-parameter teacher-KD QAT base), fine-tuned to correct and rewrite raw speech-to-text output: punctuation, capitalization, spoken-number conversion, grammar, filler removal, and full editorial rewrites, in a strict JSON contract.

This is the production quality tier. Its sibling Yooz-Light-v3-Qwen3.5-0.8B is the fast tier.

Behavior contract

Input: a system instruction plus the raw transcription as the user turn.
Output: a single JSON object, {"result": "corrected text"}, nothing else.
Non-thinking, always. The upstream Qwen3.5-4B template defaults thinking ON; this repo's template hard-pins the closed <think> block that matches the fine-tune's training rendering. Consumers do not need to pass enable_thinking=False; the flag is ignored.
Fused weights only; no LoRA adapters ship in this repo.
generation_config.json lists both stop ids (<|im_end|>, <|endoftext|>); generation terminates correctly out of the box.

Evaluation

Held-out test split (3,624 dictation-correction samples, proofread + rewrite), greedy decoding, versus the incumbent Qwen3.5-9B the correction stack used before the KD program:

Model	Exact match	Avg CER	Semantic sim	Latency*	Disk
Yooz-Quality v3 (this repo)	27.0%	0.1307	0.9442	1,226 ms	3.42 GB
Yooz-Light v3 (0.8B)	21.1%	0.1738	0.9335	341 ms	0.61 GB
Qwen3.5-9B-8bit (reference)	18.5%	0.1695	0.9363	2,427 ms	10.43 GB

*Python mlx-lm harness on an M4 Pro, batch 1; production Swift latencies are lower.

Yooz-Quality v3 beats the 9B reference on every quality metric at roughly twice the speed and a third of the disk footprint.

Usage (mlx-lm)

python

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("YoozLabs/Yooz-Quality-v3-Qwen3.5-4B")

prompt = tokenizer.apply_chat_template(
    [
        {
            "role": "system",
            "content": (
                "Rewrite this voice transcription into clear, polished "
                "text. Remove filler words, fix grammar, and convert "
                'spoken numbers to digits. Return JSON: '
                '{"result": "corrected text"}'
            ),
        },
        {
            "role": "user",
            "content": "so um i think we should you know prepare for the "
            "presentation on friday",
        },
    ],
    tokenize=False,
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512,
               sampler=make_sampler(temp=0.0)))
# {"result": "We should prepare for the presentation on Friday."}

Greedy decoding (temp=0.0) is recommended; the model is trained for deterministic correction, not creative generation.

Lineage

Base: YoozLabs/Qwen3.5-4B-qat-lean-4bit-mlx, a teacher-KD QAT full-parameter 4-bit g64 MLX build that repairs the aggregation collapse of community 4-bit PTQ at this size.
Fine-tuned (QLoRA, fused) on a proprietary gold-standard corpus of 18,285 real dictation transcriptions with two correction targets per entry (proofread and rewrite).
Ships at 6-bit g64: re-quantizing the fused LoRA delta at 4-bit on the QAT-optimized grid erases part of the fine-tune gain; 6-bit preserves adapter parity (verified on the full held-out split).
Supersedes the Yooz-Quality v2 / v2.1 working models and replaces the Qwen3.5-9B fallback in the Yooz correction stack.

Limitations

English-centric; trained on dictation/STT output, not general text.
Training sequence window 512 tokens; intended for utterance- to paragraph-length corrections.
The JSON contract is the supported interface; free-form prompting is out of scope.

Contact

dev@yooz.info

Part of the Yooz ecosystem. Sovereign Intelligence.