Skip to content
BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
README.md
README.md
 
 
README_reorganized_by_base.md
README_reorganized_by_base.md
 
 

Repository files navigation

Local LLMs for Translation

Last updated: 2026-07-13

A curated list of local and open-weight LLMs under 14B that may be useful for translation tasks.

Translation quality depends on the language pair, model size, fine-tuning data, quantization, prompt style, and the type of content being translated. The models listed here may or may not meet your expectations in every use case.

For more background, see Models & Performance.

Table of Contents

What is this?

This README is a compact collection of small and mid-sized local LLMs that can be used for translation.

The list focuses on:

  • models under 14B parameters;
  • models that are practical for local deployment;
  • models with GGUF availability when possible;
  • models with community attention, downloads, likes, language-specific fine-tuning, or reported translation usefulness.

The list is not a benchmark. It is a navigation page for finding models worth testing.

How to Choose a Model

A simple selection path:

  1. Check your available VRAM first. Local model choice is usually limited by GPU memory.
  2. Choose by language pair. A smaller model fine-tuned for your language pair may beat a larger general model.
  3. Prefer tested models marked with ★. The mark means the model has been tested on some languages or tasks.
  4. Prefer GGUF if you use llama.cpp-based tools. This includes LM Studio, Ollama, Kobold, and similar local frontends.
  5. Use stronger quantization if VRAM allows. Q5_K and above usually have no noticeable quality loss for many use cases.
  6. Test with your own content. Translation quality varies heavily across novels, subtitles, technical text, UI strings, casual chat, and document translation.

Read How to choose a LLM-Local models for more details

VRAM Tier Required VRAM Note
0~3B Vram≤4GB low vram/cpu
4B/5B Vram≈4GB~6GB Small local models/cpu
7B-10B Vram≈6G~8G Usually decent translation quality
12B-14B Vram≈12G~16G Larger local models around 14B±

Legend

Term / Symbol Meaning
★ Tested on some languages or tasks
GGUF Common model format for llama.cpp, LM Studio, Ollama, Kobold, etc.
General multilingual Broad multilingual capability, not necessarily specialized for one language pair
EN ⇌ XX English to/from multiple languages
JP Japanese
ZH Chinese
KO Korean
RU Russian
FR French
ES Spanish
UKR Ukrainian
BG Bulgarian
TR Turkish
SEA Southeast Asian languages
ne Nepali
Abliterated / uncensored / heretic / amoral / evil / ara Safety-reduced fine-tune styles. Use with caution and test output quality yourself.

Essential Reading

Read these first if you are new to local LLMs:

Heretic Grimoire Search, filter, collect, and download reproducibility records from public Heretic model archives

Model Families

Models below include base models and fine-tunes from Qwen, Gemma, Mistral, Hunyuan, and other families.

How to Read Model Names

Example: Qwen3-8B-Thinking-2507-abliterated-Q8_0-gguf

This model name is hypothetical and only used for illustration.

  • Qwen3: base model family.
  • 8B: parameter count. Common sizes include 4B, 8B, 14B, 30B, 70B, and larger.
  • Thinking: indicates a thinking-mode model. Not all thinking models are explicitly labeled.
  • 2507: time tag, usually meaning 2025/07 or a similar release/update marker.
  • abliterated: a safety-reduced fine-tune style. Similar terms include uncensored, NSFW, heretic, amoral, and evil.
  • Q8_0 / GGUF: llama.cpp quantization type and file format.

For VRAM and parameter-size guidance, see the LLM VRAM usage lists.

General Notes

Notes

  1. Quantizations from Q5_K and above usually have no noticeable quality loss in many practical cases.
  2. Larger model size × better data × more compute usually means better performance, but language-specific fine-tuning can change the result.
    • Example: Qwen3-4B < Qwen3-8B < Qwen3-14B < Qwen3-32B.
  3. Translation quality also depends on the quality and amount of training data for the target language pair.

Download the GGUF file unless you are not using a llama.cpp-based backend, such as LM Studio, Ollama, Kobold, or similar tools.

For Mac with Apple silicon, see LM Studio with Apple MLX.

Model List by VRAM

0–3B / VRAM ≤ 4GB

LLMs at this scale usually have limited translation quality.
All links are from Hugging Face.
Models are selected based on community reports, downloads, likes, and fine-tuning languages.

Model Name Languages Base Model Family Link(GGUF) Link(Original Page)
shisa-ai/shisa-v2.1-lfm2-1.2b
shisa-ai/shisa-v2.1-llama3.2-3b
LiquidAI/LFM2.5-1.2B-JP-202606		 JP ⇌ EN LFM2
Llama3.2
LFM2.5		 GGUF
GGUF
GGUF		 Original
Original
Original
LGAI-EXAONE/EXAONE-4.0-1.2B ES ⇌ EN ⇌ KO EXAONE4.0 GGUF Original
tencent/HY-MT1.5-1.8B
★tencent/HY-MT2-1.8B		 General multilingual
(33 languages)		 HY-MT1.5
HY-MT2		 GGUF
GGUF		 Original
Original
Unbabel/Tower-Plus-2B General Gemma2 GGUF Original
prithivMLmods/Qwen3-VL-2B-Instruct-abliterated-v1 EN ⇌ XX 1 Qwen3VL GGUF Original
Goekdeniz-Guelmez/Josiefied-Qwen3-1.7B-abliterated-v1 General Qwen3 GGUF Original
mistralai/Ministral-3-3B-Instruct-2512 General Mistral3 GGUF Original
tiiuae/Falcon-H1-1.5B-Deep-Instruct
tiiuae/Falcon-H1-3B-Instruct		 General FalconH1 GGUF
GGUF		 Original
Original
CohereLabs/tiny-aya series (3b) General aya Same link Original Collection

4B–5B / VRAM ≈ 4–6GB

★ means model has been tested on some languages/tasks.

Model Name Languages Base Model Family Link(GGUF) Link(Original Page)
sarvamai/sarvam-translate (note: 4B) 22 official Indian languages2 Gemma3 GGUF Original
INSAIT-Institute/MamayLM-Gemma-3-4B-IT-v1.0 UKR ⇌ EN Gemma3 GGUF Original
INSAIT-Institute/BgGPT-Gemma-3-4B-IT BG ⇌ EN Gemma3 GGUF Original
google/translategemma-4b-it General Gemma3 GGUF Original
★mlabonne/gemma-3-4b-it-abliterated-v2 General Gemma3 GGUF Original
p-e-w/gemma-4-E2B-it-heretic-ara General Gemma4 GGUF Original
himalaya-ai/himalaya-gemma-4-e2b-it NE ⇌ EN Gemma4 GGUF Original
★Goekdeniz-Guelmez/Josiefied-Qwen3-4B-Instruct-2507-gabliterated-v2 General Qwen3 GGUF Original
RefalMachine/RuadaptQwen3-4B-Instruct RU ⇌ EN Qwen3 GGUF Original
★SakuraLLM/GalTransl-v4-4B-2601 JP ⇌ ZH Qwen3 GGUF Original
★DavidAU/Qwen3-4B-2507-Thinking-heretic-abliterated-uncensored ([!]long thinking3) General Qwen3 GGUF Original
prithivMLmods/Qwen3-VL-4B-Instruct-abliterated-v1 EN ⇌ XX Qwen3VL GGUF Original
aisingapore/Qwen-SEA-LION-v4-4B-VL
aisingapore/Gemma-SEA-LION-v4.5-E2B-IT		 EN + 7 key SEA languages4 Qwen3VL
Gemma 4		 GGUF
GGUF		 Original
Original
HiTZ/Latxa-Qwen3-VL-4B-Instruct
HiTZ/Latxa-Qwen3.5-4B		 (Basque-adapted) 5+5 Qwen3VL GGUF
N/A		 Original
Original
MuXodious/gemma-3n-E2B-it-absolute-heresy (5B / CPU-optimized)6 General Gemma3n GGUF Original

7B–10B / VRAM ≈ 6–8GB

New LLMs at this size usually provide decent translation quality.

Model Name Languages Base Model Family Link(GGUF) Link(Original Page)
tiiuae/Falcon-H1-7B-Instruct General FalconH1 GGUF Original
★tencent/Hunyuan-MT2-7B General multilingual (33 languages) HY-MT2 GGUF Original
★MuXodious/gemma-3n-E4B-it-absolute-heresy (7B/CPU-optimized) General Gemma3n GGUF Original
★google/gemma-4-E4B-it (8B / CPU-optimized / MTP speedup) General Gemma4 GGUF Original
MTP
★Goekdeniz-Guelmez/Josiefied-Qwen3-8B-abliterated-v1 General Qwen3 GGUF Original
★mlabonne/Qwen3-8B-abliterated General Qwen3 GGUF Original
AvitoTech/avibe (8B)
t-tech/T-lite-it-2.1 (8B)
RefalMachine/RuadaptQwen3-8B-Hybrid		 RU ⇌ EN Qwen3 GGUF
GGUF
GGUF		 Original
Original
Original
legmlai/legml-v1.0-8b-instruct FR ⇌ EN Qwen3 GGUF Original
★shisa-ai/shisa-v2.1-qwen3-8b
nvidia/NVIDIA-Nemotron-Nano-9B-v2-Japanese		 JP ⇌ EN Qwen3
Nemotron 2 Nano		 GGUF
GGUF		 Original
Original
★prithivMLmods/Qwen3-VL-8B-Instruct-abliterated-v2 EN ⇌ XX Qwen3VL GGUF Original
aisingapore/Qwen-SEA-LION-v4-8B-VL EN + 7 key SEA languages4 Qwen3VL GGUF Original
★SakuraLLM/Sakura-GalTransl-7B-v3.7 JP ⇌ ZH Qwen2.5 GGUF Original
★mistralai/Ministral-3-8B-Instruct-2512 General Ministral3 GGUF Original
lenML/aya-expanse-8b-abliterated General aya23 GGUF Original
★Unbabel/Tower-Plus-9B General Gemma2 GGUF Original
KORMo-Team/KORMo-10B-sft
LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct
skt/A.X-4.0-Light		 KO ⇌ EN KORMo
EXAONE-3.5
Qwen2.5		 GGUF
GGUF
GGUF		 Original
Original
Original

12B–14B / VRAM ≈ 12–16GB

Model Name Languages Base Model Family Link(GGUF) Link(Original Page)
★Goekdeniz-Guelmez/Josiefied-Qwen3-14B-abliterated-v3 General Qwen3 GGUF Original
★mlabonne/Qwen3-14B-abliterated General Qwen3 GGUF Original
NousResearch/Hermes-4-14B General Qwen3 GGUF Original
Sunbird/Sunflower-14B Ugandan languages7 ⇌ EN Qwen3 GGUF Original
suayptalha/Sungur-14B TR ⇌ EN Qwen3 GGUF Original
sugoitoolkit/Sugoi-14B-Ultra-HF JP ⇌ EN Qwen2.5 GGUF Original
★SakuraLLM/Sakura-GalTransl-14B-v3.8 JP ⇌ ZH Qwen2.5 GGUF Original
google/gemma-4-12B-it
★google/gemma-4-12B-it-qat(8G VRAM support)
★coder3101/gemma-4-12B-it-qat-q4_0-unquantized-heretic(8G VRAM support)		 General multilingual (except Chinese) Gemma4 GGUF
GGUF
GGUF		 Original
Original
Original
★google/translategemma-12b-it General Gemma3 GGUF Original
INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0
INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v2.0		 UKR ⇌ EN Gemma3 GGUF
GGUF		 Original
Original
INSAIT-Institute/BgGPT-Gemma-3-12B-IT BG ⇌ EN Gemma3 GGUF Original
★mlabonne/gemma-3-12b-it-abliterated-v2
mlabonne/gemma-3-12b-it-qat-abliterated		 General Gemma3 GGUF
GGUF		 Original
Original
★mistralai/Ministral-3-14B-Instruct-2512 General Ministral GGUF Original
jenerallee78/Ministral-3-14B-abliterated General Ministral GGUF Original
mistralai/Mistral-Nemo-Instruct-2407 General Mistralnemo GGUF Original
Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24 RU ⇌ EN Mistralnemo GGUF Original
★shisa-ai/shisa-v2-mistral-nemo-12b JP ⇌ EN Mistralnemo GGUF Original

Language Index

Use this section as a second entry point if you already know your target language pair.

Japanese ⇌ Chinese

  • SakuraLLM/GalTransl-v4-4B-2601
  • SakuraLLM/Sakura-GalTransl-7B-v3.7
  • SakuraLLM/Sakura-GalTransl-14B-v3.8

Japanese ⇌ English

  • shisa-ai/shisa-v2.1-lfm2-1.2b
  • shisa-ai/shisa-v2.1-llama3.2-3b
  • LiquidAI/LFM2.5-1.2B-JP-202606
  • shisa-ai/shisa-v2.1-qwen3-8b
  • NVIDIA-Nemotron-Nano-9B-v2-Japanese
  • sugoitoolkit/Sugoi-14B-Ultra-HF
  • shisa-ai/shisa-v2-mistral-nemo-12b

Korean ⇌ English

  • LGAI-EXAONE/EXAONE-4.0-1.2B
  • KORMo-Team/KORMo-10B-sft
  • LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct
  • skt/A.X-4.0-Light

Russian ⇌ English

  • RefalMachine/RuadaptQwen3-4B-Instruct
  • AvitoTech/avibe
  • t-tech/T-lite-it-2.1
  • RefalMachine/RuadaptQwen3-8B-Hybrid
  • Vikhrmodels/Vikhr-Nemo-12B-Instruct-R-21-09-24

Ukrainian / Bulgarian / Turkish ⇌ English

  • INSAIT-Institute/MamayLM-Gemma-3-4B-IT-v1.0
  • INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0
  • INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v2.0
  • INSAIT-Institute/BgGPT-Gemma-3-4B-IT
  • INSAIT-Institute/BgGPT-Gemma-3-12B-IT
  • suayptalha/Sungur-14B

French ⇌ English

  • legmlai/legml-v1.0-8b-instruct

Nepali ⇌ English

  • himalaya-ai/himalaya-gemma-4-e2b-it

Indian Languages

  • sarvamai/sarvam-translate

Southeast Asian Languages

  • aisingapore/Qwen-SEA-LION-v4-4B-VL
  • aisingapore/Gemma-SEA-LION-v4.5-E2B-IT
  • aisingapore/Qwen-SEA-LION-v4-8B-VL

Basque / Galician / Catalan / Spanish / English

  • HiTZ/Latxa-Qwen3-VL-4B-Instruct
  • HiTZ/Latxa-Qwen3.5-4B

Disclaimer

This list is not a formal benchmark. Model quality changes with prompts, sampling settings, quantization, backend, language pair, and content type. Always test models with your own translation samples before relying on them for serious work.

Footnotes

  1. Fine-tuned on English. The same applies below. ↩

  2. Including Assamese, Bengali, Bodo, Dogri, Gujarati, English, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu ↩

  3. This CoT is very long. ↩

  4. Including Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese ↩ ↩2

  5. Including Basque, Galician, Catalan, Spanish, English and more ↩

  6. Gemma3n series (E2B/E4B) is optimized for CPU performance. Check Gemma3n for more details. ↩

  7. 31 Ugandan languages Acoli Adhola Alur Bari Chiga Gwere Kumam Karamojong Kakwa Kinyarwanda Konzo Kupsabiny Lango (Uganda) Lugbara Saamia Aringa Ganda Ma'di Masaaba Nyole Nyankole Nyoro Pokangá Gungu Ruuli Amba (Uganda); Swahili (macrolanguage); Teso Talinga-Bwisi Tooro Soga + English ↩

About

This is only a small list of Collection for llm models under 14B which are used for translation. All models (add) support for more than 119 languages.

Topics

languagellmmultilingualtranslation

Resources

Readme
Activity

Stars

4 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors