Qwen3-1.7B-EasyLanguage (4-bit, MLX)

This model is a LoRA fine-tune of Qwen/Qwen3-1.7B. It rewrites live speech transcripts into easy-to-read registers across 29 language locales: German (Leichte Sprache), French (FALC), English (Easy/Plain English), Spanish (Lectura Fácil), Arabic, and 24 others. Each locale follows its national or European easy-language standard. We quantized the model to 4 bits for Apple-silicon inference with MLX.

The model splits run-on sentences, removes disfluencies, keeps names and numbers, and stays in the input language. We trained it on a lean prompt with no few-shot examples so it internalizes the register. Shorter prompts reduce live-caption latency.

Part of Live Linguist. The apps for macOS, iPhone/iPad, Android, and PC are pending store release; the macOS build is available now. The open models and dataset are available today.

Project · Apps · All benchmarks · Research · Dataset

The specialist case

General-purpose models give helpful answers; easy language requires a register. We trained this fine-tune against the same deterministic validators the app uses at runtime: sentence caps for each language, simple-grammar checks, source anchoring, number fidelity, anti-parroting, and a wrong-language guard. The validators make the register the objective.

Across 27 whitespace-delimited locales, the 1.7B fine-tune averages SARI 54.6 against 49.6 for stock Qwen3-1.7B and 97% register compliance against 74%. It beats the 0.6B fine-tune on SARI in all 27 locales and keeps a 99.6% in-language rate.

The 1.7B model provides the quality option for recorded material, offline work, and slower conversations. The 0.6B model drives live captions.

Evaluation (held-out test, ~200 sentences per language)

metric 1.7B fine-tuned stock Qwen3-1.7B
Mean SARI (27 whitespace locales) 54.6 49.6
Mean register compliance (29 locales) 97% 74%
Mean language ID (29 locales) 99.6% 99.4%
Mean parrot rate (29 locales) 0% 1%

Higher is better in every row except parrot rate, where lower is better.

What SARI measures. SARI compares a model's simplified sentence with a reference easy-language version. It gives credit for words the model kept, added, or deleted in the right places, and it penalizes changes that break the match. Scores run from 0 to 100, and higher is better. The number is not a percentage, and the absolute score depends on the test set, so read the gap between models on the same rows.

lang register SARI ft SARI stock chrF ft chrF stock compliance ft compliance stock LID ft parrot ft parrot stock
vi ETR / ISO plain language 62.5 57.3 60.9 54.6 86% 50% 100% 0% 0%
en Easy / Plain English 62.4 52.0 57.1 42.7 100% 68% 100% 0% 1%
fr FALC 62.0 54.1 55.6 45.9 100% 61% 100% 0% 0%
es Lectura Fácil 61.7 56.9 57.8 51.4 99% 61% 100% 0% 0%
pt-PT Leitura Fácil 61.2 54.3 61.0 51.8 96% 82% 100% 0% 0%
pt-BR Leitura Fácil 57.8 56.9 54.0 49.7 89% 32% 100% 0% 0%
it Easy to Read 56.2 49.5 51.9 40.6 89% 48% 100% 0% 1%
da Letlæst 55.6 49.2 52.8 42.1 98% 80% 100% 0% 0%
nl Makkelijk Lezen 55.1 51.8 51.4 46.4 94% 44% 98% 0% 0%
sv Lättläst 55.0 49.5 57.9 48.2 96% 89% 98% 0% 0%
hi Easy-to-Read 54.5 49.2 47.6 40.2 84% 60% 98% 0% 0%
ru Easy-to-Read 54.5 51.0 51.5 43.7 97% 72% 100% 0% 0%
nb Lettlest 54.3 41.4 56.2 38.1 99% 96% 100% 0% 0%
ro Easy-to-Read 53.7 48.1 52.9 41.5 100% 92% 100% 0% 0%
hr Easy-to-Read 53.5 45.4 54.5 42.6 100% 86% 100% 0% 0%
uk Easy-to-Read 53.1 51.8 54.8 51.4 100% 88% 100% 0% 0%
bg Easy-to-Read 52.7 51.7 55.0 52.5 98% 73% 99% 0% 0%
ar Easy-to-Read (RTL) 52.1 51.0 50.5 46.3 98% 68% 100% 0% 0%
pl Easy-to-Read 51.5 44.5 49.8 36.6 100% 82% 100% 0% 0%
sk Easy-to-Read 51.5 46.9 47.7 40.3 96% 82% 100% 0% 2%
de Leichte Sprache 51.1 41.9 46.6 31.1 100% 60% 100% 0% 18%
hu Easy-to-Read 50.7 49.8 52.3 50.0 100% 82% 100% 0% 0%
fi Selkokieli 50.6 47.2 56.0 48.8 100% 91% 100% 0% 0%
tr Kolay Anlaşılır 50.6 46.4 50.6 44.3 99% 85% 100% 0% 0%
cs Snadné čtení 50.4 48.6 49.2 45.9 100% 81% 100% 0% 0%
ko Easy-to-Read 50.1 46.5 42.9 33.7 98% 88% 100% 0% 0%
et Easy-to-Read 48.4 45.6 50.7 45.5 99% 73% 100% 0% 0%
zh-CN ISO plain language n/a n/a 53.9 41.6 99% 94% 100% 0% 0%
ja Yasashii Nihongo n/a n/a 45.5 38.8 98% 92% 100% 0% 0%

Reading the table

  • SARI measures simplification quality (Xu et al., 2016): words kept, added, and deleted relative to references. chrF is a character n-gram F-score (Popović, 2015), and it helps where whitespace tokenization fails, as it does for ja and zh-CN.
  • Compliance is the share of outputs that pass the target register's deterministic validators. LID is the share that stay in the target language. Parrot is the share that copy the input or a prompt example.
  • The table shows fine-tuned and stock values in paired columns. SARI reads n/a for ja and zh-CN because whitespace tokenization makes it meaningless for those scripts; those locales use chrF, compliance, and LID.
  • The baseline is stock Qwen3 at the same size, 4-bit quantization, and prompt budget. That comparison isolates the fine-tune.
  • We leave proprietary API models out of this table. No public per-language easy-language benchmark exists for them, and we publish measured numbers. The harness works with any model, so a reproducible proprietary baseline can join the table later.

Edge profile

  • 4-bit MLX, about 934 MB.
  • Runs on the device. The model sends no inference request, transcript, or audio sample off the device.
  • We tested the macOS app on M1 (16 GB), M2 (96 GB), M3 (128 GB), and A18 Pro (8 GB).
  • We publish model size and device coverage. We do not publish tokens-per-second figures we have not verified. Latency depends on the speech segment, the speech-recognition front end, and the device.

Usage (MLX)

from mlx_lm import load, generate
model, tok = load("ndgold/Qwen3-1.7B-EasyLanguage-4bit")
msgs = [{"role":"system","content":"<framework system prompt>"},
        {"role":"user","content":"Original: <utterance>\nRewritten:"}]
p = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False, enable_thinking=False)
print(generate(model, tok, prompt=p, max_tokens=96))

Citation

If you use this model in teaching, research, or a published build, attribute Dylan Goldblatt.

APA 7

Goldblatt, D. (2026). Qwen3-1.7B-EasyLanguage-4bit [Model]. Hugging Face. https://huggingface.co/ndgold/Qwen3-1.7B-EasyLanguage-4bit

BibTeX

@misc{goldblatt2026easylanguage17b,
  author    = {Goldblatt, Dylan},
  title     = {Qwen3-1.7B-EasyLanguage-4bit},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/ndgold/Qwen3-1.7B-EasyLanguage-4bit}
}

For the project as a whole, see the citation samples at ndgold.com/live-linguist.

Educational use & ownership

Dylan Goldblatt created the models and the accompanying dataset for educational purposes. They support language learners, educators, accessibility teams, and researchers. They do not replace a certified accessibility service, a medical device, or human easy-language review.

The open models carry Apache-2.0; the dataset carries CC BY-SA 4.0. Aquilops LLC holds the sole intellectual property in the closed-source Live Linguist applications. The apps cost nothing, and their source stays closed.

Sources & licenses

  • Base model: Qwen3 (1.7B, Apache-2.0).
  • German seed data: tum-nlp/German4All-Corpus (German Wikipedia, CC BY-SA).
  • Synthetic pairs (all other languages and German augmentation): we generated spoken-to-easy-language pairs with Claude (Anthropic) for Leichte Sprache (de), FALC (fr), Easy/Plain English (en), Lectura Fácil (es), and 25 further locales that follow each language's Easy-to-Read or plain-language standard (Inclusion Europe "Information for All", Selkokieli, Lättläst, やさしい日本語, ISO 24495-1, and others). Every pair passed a deterministic validator suite: sentence-length caps for each language, language ID, fidelity anchoring, number preservation, and anti-parroting.
  • The model serves the Live Linguist on-device live-caption simplifier. It is not a general chatbot.
Downloads last month
44
Safetensors
Model size
2B params
Tensor type
U32
·
BF16
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

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

Model tree for ndgold/Qwen3-1.7B-EasyLanguage-4bit

Finetuned
Qwen/Qwen3-1.7B
Quantized
(551)
this model