Qwen3.5-0.8B GEC for Kazakh, Russian, and English

A compact 0.8B grammatical-error-correction / proofreading model for Kazakh, Russian, and English. It takes raw text โ€” voice-dictation output, chat messages, quick notes โ€” and cleans it up: restores punctuation, capitalization, and paragraph breaks, and fixes typos and spelling errors โ€” while preserving your words, meaning, language, and tone. It does not rewrite, paraphrase, translate, soften, or add anything of its own.

Published as GGUF (Q4_0) for CPU inference with llama.cpp, LM Studio, and Ollama. ~537 MB, ~44 tokens/s on CPU.

By Loqira Labs โ€” everyday AI you own, not rent. Runs on the hardware you already have โ€” no cloud, no frontier API.


โšก TL;DR

  • Task: grammar / spelling / punctuation / capitalization / paragraphing correction (GEC, proofreading).
  • Word-preserving: keeps your exact words, meaning and tone; it corrects, it does not rewrite.
  • Languages: Kazakh, Russian, English (incl. Kazakh special letters).
  • One request โ†’ one response. This is a single-turn model: use it for one-shot calls, not multi-turn chat.
  • Must-have runtime settings: the exact system prompt below, thinking OFF, temperature 0.
  • Format: GGUF Q4_0, ~537 MB, ~44 tok/s CPU.

What it does

  • Restores punctuation, capitalization, and paragraph structure.
  • Fixes typos and spelling errors.
  • Word-preserving โ€” keeps your words, meaning, tone, and language intact.
  • Adds nothing, invents nothing, and does not soften or rephrase.

โš ๏ธ Single-turn model โ€” one request, one response

This model is trained and meant to be used as a one-shot call: send one input, get one corrected output. It is not a conversational model โ€” do not build multi-turn dialogues, do not feed prior turns back in, and do not chain calls in a running chat context. Each correction is an independent request with a fresh context and the same system prompt.


Languages

  • Explicitly trained and evaluated: Kazakh (kk), Russian (ru), English (en) โ€” including Kazakh-specific letters (า›, ำฉ, าฑ, าฏ, ำ™, ั–, าฃ, า“, าป), which the model restores.
  • Other languages: the base model (Qwen3.5-0.8B) additionally covers 198+ languages. The correction behaviour may transfer to them, but this was not tested โ€” treat any non-KK/RU/EN use as experimental.

Model details

Base model Qwen/Qwen3.5-0.8B (hybrid: GatedDeltaNet + full attention, 24 layers, hidden 1024)
Relation finetune (quantization-aware)
Method QAT at 4/8/16-bit with train==deploy parity โ€” our int4/int8 codes are injected into the GGUF, not re-quantized
Format GGUF, Q4_0 weights + Q8_0 / F32 (short-conv, norms)
Size ~537 MB
Speed ~44 tokens/s on CPU
Training context 2048 tokens (โ‰ˆ1600 per sample on average)
Chat template + stop token Both baked into the GGUF

The stop token is <|im_end|> โ€” clients that read the GGUF metadata pick it up automatically.


Usage

1. System prompt โ€” required (Russian, verbatim)

The model was trained to always receive this exact system prompt. Pass it as the system message. It is a single unified prompt โ€” the same for every input and every language (Kazakh, Russian, and English input all use this one prompt).

ะžั‚ั„ะพั€ะผะฐั‚ะธั€ัƒะน ั‚ะตะบัั‚ ะณะพะปะพัะพะฒะพะณะพ ะฒะฒะพะดะฐ: ั€ะฐััั‚ะฐะฒัŒ ะฟัƒะฝะบั‚ัƒะฐั†ะธัŽ ะธ ะทะฐะณะปะฐะฒะฝั‹ะต ะฑัƒะบะฒั‹, ั€ะฐะทะฑะตะน ะฝะฐ ะฐะฑะทะฐั†ั‹, ะธัะฟั€ะฐะฒัŒ ะพะฟะตั‡ะฐั‚ะบะธ ะธ ะพั€ั„ะพะณั€ะฐั„ะธั‡ะตัะบะธะต ะพัˆะธะฑะบะธ. ะกะพั…ั€ะฐะฝะธ ัะทั‹ะบ, ัะปะพะฒะฐ ะธ ัะผั‹ัะป, ะฝะธั‡ะตะณะพ ะฝะต ะดะพะฑะฐะฒะปัะน ะพั‚ ัะตะฑั.

The prompt is in Russian by design โ€” that is what the model was trained on, so it gives the best results. A version of the prompt translated into another language also works, but noticeably worse โ€” keep the Russian prompt unless you have a specific reason not to. Without a system prompt at all (or with an unrelated one) the model is out-of-distribution and quality drops sharply.

2. Thinking OFF โ€” required

The model was trained with an empty reasoning block (<think>\n\n</think>\n\n). A client with "thinking / reasoning" enabled will inject <think>\n and push the model out-of-distribution (garbled prefix, quality loss). Keep thinking OFF and use temperature 0 (greedy, deterministic).

3. llama.cpp (llama-cli)

llama-cli -m Qwen3.5-0.8B-GEC-KAZ-RUS-ENG.Q4_0.gguf \
  -sys "ะžั‚ั„ะพั€ะผะฐั‚ะธั€ัƒะน ั‚ะตะบัั‚ ะณะพะปะพัะพะฒะพะณะพ ะฒะฒะพะดะฐ: ั€ะฐััั‚ะฐะฒัŒ ะฟัƒะฝะบั‚ัƒะฐั†ะธัŽ ะธ ะทะฐะณะปะฐะฒะฝั‹ะต ะฑัƒะบะฒั‹, ั€ะฐะทะฑะตะน ะฝะฐ ะฐะฑะทะฐั†ั‹, ะธัะฟั€ะฐะฒัŒ ะพะฟะตั‡ะฐั‚ะบะธ ะธ ะพั€ั„ะพะณั€ะฐั„ะธั‡ะตัะบะธะต ะพัˆะธะฑะบะธ. ะกะพั…ั€ะฐะฝะธ ัะทั‹ะบ, ัะปะพะฒะฐ ะธ ัะผั‹ัะป, ะฝะธั‡ะตะณะพ ะฝะต ะดะพะฑะฐะฒะปัะน ะพั‚ ัะตะฑั." \
  -p "ะฟั€ะธะฒั‚ะต ะฒัะตะผ ัะตะณะดะฝั ั ั…ะพั‡ัŽ ั€ะฐัะบะฐะทะฐั‚ัŒ ะฟั€ะพ ะฝะฐัˆ ะฝะพะฒั‹ะน ะฟั€ะฐะตะบั‚ ะฝะฐะด ะบะพั‚ั€ั‹ะผ ะผั‹ ั€ะฐะฑะพั‚ะฐะตะผ ัƒะถะต ะฟะพั‡ั‚ะธ ะฟะพะป ะณะพะดะฐ" \
  -st -no-cnv --reasoning off --reasoning-format none \
  --chat-template-kwargs "{\"enable_thinking\": false}" \
  --temp 0 --top-k 1 --repeat-penalty 1.0 -n 512

For long inputs, raise -n (e.g. 1024โ€“2048) so the output isn't truncated.

Keep each request short. The model was trained on a 2048-token context window (โ‰ˆ1600 tokens per sample on average). It is a single-shot formatter, not a long-document processor โ€” feeding a much longer sequence in one call pushes it out-of-distribution and degrades quality. Split long text into chunks (e.g. by paragraph) and format each chunk in its own request, keeping prompt + output within ~2048 tokens.

4. llama-server (OpenAI-compatible API)

Start the server with --reasoning-format none. In each request send the system prompt above, temperature: 0, and "chat_template_kwargs": {"enable_thinking": false}. Strip any leading <think>...</think> from the returned content.

5. Ollama

A ready Modelfile is included โ€” it bakes in the correct chat template (think-off), the system prompt, and greedy decoding.

ollama create qwen3.5-gec -f Modelfile
ollama run qwen3.5-gec "ะฟั€ะธะฒั‚ะต ะฒัะตะผ ัะตะณะดะฝั ั ั…ะพั‡ัŽ ั€ะฐัะบะฐะทะฐั‚ัŒ ะฟั€ะพ ะฝะฐัˆ ะฝะพะฒั‹ะน ะฟั€ะฐะตะบั‚"

6. LM Studio

Load the GGUF, disable "thinking / reasoning", set temperature 0, and set the system prompt above. Verify that the rendered template produces <think>\n\n</think>\n\n (an empty block), not <think>\n.

Compatibility

This model uses the Qwen3.5 architecture (qwen35 โ€” hybrid GatedDeltaNet + attention), supported in mainline llama.cpp and the apps built on it. Just use a recent llama.cpp / LM Studio / Ollama. If you ever see unknown model architecture 'qwen35', your app bundles an outdated llama.cpp โ€” update it. Verified on the CPU path; on a GPU backend, run the golden-probe examples below first.


Examples (temperature 0, thinking off)

Russian โ€” typos:

in : ะฟั€ะธะฒั‚ะต ะฒัะตะผ ัะตะณะดะฝั ั ั…ะพั‡ัŽ ั€ะฐัะบะฐะทะฐั‚ัŒ ะฟั€ะพ ะฝะฐัˆ ะฝะพะฒั‹ะน ะฟั€ะฐะตะบั‚ ะฝะฐะด ะบะพั‚ั€ั‹ะผ ะผั‹ ั€ะฐะฑะพั‚ะฐะตะผ ัƒะถะต ะฟะพั‡ั‚ะธ ะฟะพะป ะณะพะดะฐ
out: ะŸั€ะธะฒะตั‚ ะฒัะตะผ! ะกะตะณะพะดะฝั ั ั…ะพั‡ัƒ ั€ะฐััะบะฐะทะฐั‚ัŒ ะฟั€ะพ ะฝะฐัˆ ะฝะพะฒั‹ะน ะฟั€ะพะตะบั‚, ะฝะฐะด ะบะพั‚ะพั€ั‹ะผ ะผั‹ ั€ะฐะฑะพั‚ะฐะตะผ ัƒะถะต ะฟะพั‡ั‚ะธ ะฟะพะป ะณะพะดะฐ.

English โ€” typos:

in : helo evryone i jsut wanted to tel you abuot the meting we have tomorow mornig at nine oclock
out: Hello everyone, I just wanted to tell you about the meeting we have tomorrow morning at nine o'clock.

Kazakh โ€” special letters restored (า› / ำฉ / ั–):

in : ะผะตะฝ ะบะฐะปะฐะณะฐ ะฑะฐั€ะดั‹ะผ ะพะฝะดะฐ ะบะพะฟ ะฐะดะฐะผ ะฑะฐั€ ะตะดะธ ะบั‹ั ะบะตะทะธ ััƒั‹ะบ ะฑะพะปะดั‹
out: ะœะตะฝ า›ะฐะปะฐา“ะฐ ะฑะฐั€ะดั‹ะผ, ะžะฝะดะฐ ะบำฉะฟ ะฐะดะฐะผ ะฑะฐั€, ะ•ะฝะดั– า›ั‹ั ะบะตะทั– ััƒั‹า› ะฑะพะปะดั‹.

Preserves tone โ€” does not soften or rewrite:

in : ะดะฐ ั‚ั‹ ะธะทะดะตะฒะฐะตัˆัŒัั ั‡ั‚ะพ ะปะธ ะพะฟัั‚ัŒ ะฒัั‘ ัะปะพะผะฐะป ั ะถะต ะณะพะฒะพั€ะธะป ะฝะต ั‚ั€ะพะณะฐะน ัั‚ะพั‚ ะบะพะด
out: ะ”ะฐ ั‚ั‹ ะธะทะดะตะฒะฐะตัˆัŒัั ั‡ั‚ะพ ะปะธ? ะžะฟัั‚ัŒ ะฒัั‘ ัะปะพะผะฐะป. ะฏ ะถะต ะณะพะฒะพั€ะธะป: ยซะะต ั‚ั€ะพะณะฐะน ัั‚ะพั‚ ะบะพะดยป.

You can use these four as a quick golden probe: if your client produces materially different output at temperature 0, the prompt or think-off setting is wrong โ€” check the system prompt and thinking-off first.


Known limitations

  • Mixed script within one word (e.g. Cyrillic + Latin "ะฑะปin") is not corrected.
  • Segmentation around URLs / @handles can be unstable (though URLs and handles themselves are preserved byte-for-byte).
  • "HH MM" time without a separator may get an extra dot.
  • ~10% of isolated in-word typos may remain (not systematic); rare Kazakh word substitutions.
  • Short context (2048 tokens, ~1600 avg per sample). Process long text in chunks โ€” one request per chunk โ€” rather than a single large sequence; see Usage.
  • 0.8B / 4-bit ceiling โ€” this is a small, quantized model built for speed and privacy, not a frontier LLM.

License

Apache-2.0. Base model: Qwen/Qwen3.5-0.8B.


About Loqira Labs

Loqira Labs builds capable AI that runs on the hardware you already have โ€” no cloud, no frontier API, no vendor to answer to. Everyday AI you own, not rent.

๐ŸŒ loqira.tech

Downloads last month
398
GGUF
Model size
0.8B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for loqira/Qwen3.5-0.8B-GEC-KAZ-RUS-ENG

Finetuned
(915)
this model