Special-purpose model. This checkpoint does exactly one job: cleaning up speech-to-text output under the fixed system prompts and JSON output contract documented below. It is not a general-purpose chat or instruction model; off-task use (open-ended chat, question answering) is unsupported and will underperform the base model. All Yooz task models are collected in Yooz Working Models; for general-purpose local LLMs see Lean QAT MLX.

Speech-to-text touchup model from Yooz Labs. Takes raw STT output and returns a cleaned-up version, either lightly proofread or more aggressively rewritten, as a single JSON object: {"result": "..."}.

This is the Light tier of the Yooz Engine touchup stack — the fast-path companion to Yooz-Quality-v2. It runs fully on-device on Apple Silicon via MLX, in the privacy-first Yooz tradition: no cloud, no logging, no leak.

What it is

Task: post-process automatic speech recognition output. Two modes:
- Proofread — minimal edits: spelling, punctuation, capitalization, obvious word-choice errors. Filler words preserved.
- Rewrite — larger edits: removes filler, fixes grammar, splits run-on sentences, normalizes spoken numbers. Meaning preserved.
Output contract: always a single JSON object of the form {"result": "<cleaned text>"}. No explanation, no markdown.
Footprint: 4-bit MLX quantized fused weights, ~265 MB on disk.
Latency target: ~200 ms per touchup on Apple Silicon.

Lineage

Base: mlx-community/Qwen2.5-0.5B-Instruct-4bit (Apache 2.0).
Output format: standalone 4-bit fused checkpoint. Unlike Quality v2 (which publishes both an adapter and the fused weights), this variant publishes the FUSED checkpoint only — avoids adapter-loader ambiguity on mlx-swift-lm consumers.
Why "v2": v1 of the Yooz Engine Light tier was an earlier Qwen2.5 fine-tune. v2 uses the canonical Qwen2.5-0.5B-Instruct-4bit base shared across the Yooz ecosystem.

Why Light vs Quality

Light is optimized for the fast, low-latency path — short utterances and mechanical fixes where the user wants surface cleanup without semantic rewriting. Use Quality for longer dictation where you want intent-preserving rewrites, self-correction handling, and tone preservation.

Use it

Pull via mlx-lm:

python

from mlx_lm import load, generate

model, tokenizer = load("YoozLabs/Yooz-Light-v2-Qwen2.5-0.5B-LoRA")

system_prompt = """Fix grammar, capitalize properly, and convert spoken numbers
to digits. Keep ALL sentences. Return the fixed text as JSON.

Always respond with ONLY a JSON object. Never remove sentences. Never include
explanations."""

raw = "the meeting is at two pm on march fifteenth"

response = generate(
    model, tokenizer,
    prompt=tokenizer.apply_chat_template(
        [{"role": "system", "content": system_prompt},
         {"role": "user", "content": raw}],
        tokenize=False, add_generation_prompt=True,
    ),
    max_tokens=256,
)
# {"result": "The meeting is at 2 PM on March 15th."}

License

Apache 2.0 (same as the base model).

YoozLabs/Yooz-Quality-v2-Qwen3.5-0.8B-LoRA — the Quality tier sibling.
yooz-engine — runtime that consumes these checkpoints.
Yooz Working Models — all Yooz task-specific models and artifacts.

Contact

Questions: dev@yooz.info
Bug reports: github.com/yooz-labs/yooz-engine

YoozLabs/Yooz-Light-v2-Qwen2.5-0.5B-LoRA

Yooz-Light-v2-Qwen2.5-0.5B-LoRA