TranslatePsy-Nano: Compact Multilingual Machine Translation for Resource-Constrained Edge Deployment

Supporting multilingual translation in a single compact model

Key Highlights

Tether Data, S.A. de C.V. (Tether AI Research, we, us, our) introduces compact multilingual machine translation models for European-English and African-English language pairs. It ships two models' families: TranslatePsy-EuroNano which consists of nine European languages—German, Spanish, French, Italian, Portuguese, Czech, Finnish, Dutch, and Swedish, and TranslatePsy-AfriNano which consists of eight African languages—Amaharic, hausa, Igbo, Lingala, Somali, Swahili, Yoruba, Zulu. Instead of maintaining separate bilingual models for each language pair, users can deploy a single checkpoint and select the target language using lightweight language tags. Each language group has three model variants: Base (42MB), Base-Memory (31MB) and Tiny (17MB) to simplify multilingual deployment on resource-constrained edge devices.

• Up to 56.7× Smaller Than Large Multilingual Models: Our largest quantized model occupies only 42 MB, making it practical for browser, mobile, desktop, and edge-device deployment. Compared to NLLB-200-600M, which requires approximately 2.4 GB of deployment storage, our model reduces deployment size by 56.7× while maintaining competitive translation quality.

• Retains 98.4% of NLLB-200 Translation Quality: Despite being dramatically smaller, the proposed multilingual model achieves an average COMET score of 0.860 on European-to-English translation, retaining approximately 98.4% of the quality achieved by NLLB-200. This demonstrates that carefully designed multilingual models can preserve most of the quality benefits of large-scale systems while drastically reducing deployment requirements.

• Optimized for CPU and Mobile Inference: Unlike many modern multilingual models that are primarily designed for GPU inference, our models are optimized for CPU deployment using the Marian/Bergamot framework. Experiments show substantial advantages in throughput, runtime memory, and mobile-device deployment, making the models suitable for offline translation scenarios.

• Adequacy-Based Data Filtering: Adequacy filtering is used to retain high-quality parallel sentence pairs that preserve the intended meaning across source and target languages. Each candidate pair is scored for translation adequacy, and low-scoring or potentially misaligned pairs are removed before training. This reduces noisy supervision and improves the overall quality and reliability of the translation training data.

• One Download Instead of many: Traditional on-device multilingual deployment often requires maintaining multiple bilingual models. Our approach replaces multiple bilingual translation models with a single multilingual checkpoint, eliminating repeated vocabularies, duplicated runtime artifacts, model switching overhead, and complex update pipelines.

---

🚀 TranslatePsy-Nano on Huggingface and Github

Copyright Complaints: We will take appropriate actions in response to notice of copyright infringement. If you believe your work has been used or copied in a manner that infringes upon your intellectual property rights, please email data-apps@tether.io identifying and describing both the copyrighted work and alleged infringing content to file a notice of infringement.

TranslatePsy-Nano CPU Performance

Following inference benchmarks are run on CPU. Server measurements were run on a Linux server with 2X Intel Xeon Platinum 8480+ CPUs, 2 TB RAM, Ubuntu 22.04, and Linux 6.8. All CPU systems use four threads and beam size 4. We report throughput, time-to-first-token, peak RAM, and model storage under this CPU deployment setting. Android/mobile results are covered later in the Mobile Deployment section.

---

Introduction

Recent advances in multilingual neural machine translation have led to the development of increasingly large models capable of supporting hundreds of languages within a single system. Models such as NLLB-200 [1], M2M-100 [2], and MADLAD-400 [3] demonstrate that multilingual parameter sharing [4] can achieve strong translation quality across a broad range of language pairs.

While these models have significantly improved multilingual coverage, their deployment requirements remain substantial. Large multilingual systems typically require hundreds of megabytes or even gigabytes of storage, high runtime memory consumption, and considerable computational resources. These requirements are often acceptable in cloud-based environments but become challenging for browser-based applications, mobile devices, enterprise desktops, and edge deployments where storage and memory budgets are limited.

An alternative deployment strategy is used by browser translation systems such as Firefox Translations [5], which employ collections of bilingual models. Although bilingual models are efficient and provide strong translation quality, supporting multiple languages requires maintaining a separate model for every language pair. As language coverage expands, deployment size, update complexity, and maintenance costs increase almost linearly.

This motivates an important question:

Can multilingual parameter sharing be used primarily as a deployment optimization rather than a language coverage optimization?

Instead of building a single model for hundreds of languages, we investigate language-group-specific multilingual translation models. The core idea is to combine related languages into a compact multilingual model that preserves the advantages of multilingual learning while maintaining a deployment footprint suitable for browser, mobile, and edge-device environments.

To evaluate this idea, we train multilingual European↔English translation models supporting nine European languages within a single checkpoint. We compare these models against strong deployment-oriented baselines, including Firefox bilingual models and NLLB-200.

Our contributions can be summarized as follows:

• We develop compact multilingual European↔English translation models optimized for deployment.
• We investigate multiple model sizes ranging from 17 MB to 42 MB.
• We evaluate translation quality using COMET across all supported languages.
• We perform deployment-oriented evaluation covering storage footprint, memory consumption, CPU throughput, and Android deployment.
• We demonstrate that compact multilingual models can retain nearly all of the translation quality of much larger multilingual systems while dramatically reducing deployment costs.

---

Methodology

Design Philosophy

The primary goal of this work is not to maximize language coverage but to maximize deployment efficiency. We therefore focus on a carefully selected language group rather than attempting to support hundreds of languages simultaneously.

The proposed multilingual models consolidate nine European languages into a single translation system. Translation direction is controlled using language-specific tags appended to the source sentence. Example: ##FR This is a multilingual translation model.

Architecture

The models are built using Marian NMT [6] and the Bergamot deployment framework [7].

Unlike standard Transformer [9] architectures, Bergamot employs a hybrid recurrent-attention architecture specifically optimized for CPU inference. The encoder consists of 6 bidirectional layers, Multi-head self-attention, Residual connections and Layer normalization. The bidirectional encoder captures contextual information from both directions of the source sentence while maintaining efficient memory usage. The decoder employs 2 SSRU layers, Cross-attention and Feed-forward networks. SSRU (Simple Recurrent Unit) [8] replaces Transformer self-attention with a lightweight recurrent mechanism. This significantly reduces decoding complexity and improves inference speed, particularly on CPUs.

Model Variants and Differences

To support different deployment requirements, we train three model configurations. The Tiny model targets highly constrained environments such as browsers and low-memory mobile devices. Base-Memory provides a middle ground between quality and efficiency, while Base maximizes translation quality while remaining deployment-friendly. These variants allow us to study the quality-efficiency trade-off under different deployment constraints.

---

Datasets

Training Data

Training data was collected from multiple publicly available multilingual resources, including EuroParl, Tatoeba, ParaCrawl, WikiMedia, OPUS-100, TED2020, News-commentary, Gobal voices, MADLAD-400.

Data Cleaning

These resources provide complementary coverage across parliamentary proceedings, legal documents, web-mined text, encyclopedic content, and multilingual web corpora. Rather than maximizing training volume, we focus on corpus quality and multilingual consistency. All datasets are processed through a unified filtering pipeline which includes 1. dropping a pair if either side has fewer than two tokens or more than 120, exceeds 1000 characters, 2. has a token-count ratio above 2.2 in either direction, 3. drop copies (identical after case/punctuation normalisation) and rows with alphabetic-character ratio below 0.5 (numbers, markup etc.).

To remove sentence pairs written in the wrong language, we apply fastText's lid.176 language identification model [10] to both sides of every pair. The predicted language is then compared with the expected language for that corpus (e.g., German text in an en--de corpus). Because language identification is less reliable for short segments, we use different acceptance criteria based on sentence length. For sentences longer than five words, the predicted language must match the expected language with confidence of at least $0.8$. Confidence score is used to see how much confident the language predictor is. We also perform adequacy score which helps to extract pairs which are mutually consistent and correct translation of each other. The objective was to maximize data quality while preserving multilingual diversity. After preprocessing and filtering, the final multilingual corpus contains approximately:

Validation and evaluation dataset

FLORES-200 dev data is used for early stopping, model selection and hyperparameter tuning. Final evaluation is performed on FLORES-200 devtest. FLORES provides a standardized multilingual benchmark with professionally curated translations across a large set of languages, making it suitable for multilingual evaluation.

---

Training Setup

All models were trained using Marian NMT.

Optimization

The following optimization settings were used:

Vocabulary

A shared SentencePiece vocabulary containing 32k subword units was trained across all languages. The shared vocabulary encourages multilingual parameter sharing and reduces deployment redundancy.

---

Benchmarks

We evaluate Quality using two metrics i.e. COMET [11] and SacreBleu [12].

European → English

The multilingual models achieve strong performance across all supported languages.

The proposed model retains approximately 98.4% of NLLB-200 quality despite being dramatically smaller.

English → European

This direction remains more challenging because the model must generate multiple target languages using a shared decoder. Nevertheless, the multilingual models maintain competitive quality while significantly reducing deployment complexity.

Generalizing the Approach: TranslatePsy-AfriNano

The European experiments above establish that a single language-group model can replace many bilingual checkpoints while staying small enough for browser and mobile deployment. TranslatePsy-AfriNano applies the same design philosophy to African↔English translation. We train multilingual model for English↔8 African languages (Amharic, Hausa, Igbo, Lingala, Somali, Swahili, Yoruba ad Zulu). We follows same architecture and training setup for African↔English translation except dataset. OPUS-100 (Parallel dataset) and MADLAD-400 datasets are used for training. MADLAD-400 has no parallel data therefore we synthetically generated back-translated data using the NLLB-200-3.3B model. The African languages sentences are translated into English using the NLLB model. Flores-200 dev and devtest are used as validation and evaluation set respectively. After multilingual training we got achieved average COMET scores as follows:

---

Deployment-Centric Evaluation

Translation quality alone does not determine practical usefulness. Deployment characteristics become increasingly important in browser, mobile, and edge-device environments. INT8 quantization introduces only minimal quality degradation. Across all model variants, average COMET reductions remain within a few thousandths of a point, demonstrating that deployment-friendly quantization can be applied without meaningful translation quality loss.

Figure 1: Throughput–Bundle Frontier

Figure 2: Quality–Size–Throughput Frontier

Storage

The multilingual approach dramatically reduces deployment footprint by consolidating multiple language pairs into a single model.

Runtime Memory

Compact multilingual models remain practical even under memory-constrained deployment scenarios.

CPU Throughput

The proposed models achieve strong throughput while maintaining competitive quality.

Mobile Deployment

Experiments on Android devices show that multilingual deployment remains practical even in offline mobile settings. Mobile measurements were run on a OnePlus 13 with Snapdragon 8 Elite, 24 GB RAM, Android 16, and a 256 MB app heap budget. Ours and Firefox both run through the @qvac/sdk [13] Bergamot path.

---

Conclusion

We presented a family of compact multilingual machine translation models designed specifically for deployment-oriented environments. By consolidating multiple languages into a single multilingual checkpoint, the proposed approach reduces deployment complexity while preserving strong translation quality.

The resulting models achieve up to 98.4% of the translation quality of NLLB-200 while reducing deployment size by 56.7× and significantly lowering runtime memory requirements. Through comprehensive evaluation covering translation quality, throughput, memory consumption, and mobile deployment, we demonstrate that language-group-specific multilingual models provide an effective balance between multilingual coverage, translation quality, and deployment efficiency.

As multilingual translation increasingly moves toward browsers, mobile devices, and edge environments, deployment-aware model design will become an increasingly important complement to traditional quality-focused evaluation.

References

[1] Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.

[2] Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, and Armand Joulin. 2020. Beyond English-centric multilingual machine translation. arXiv preprint arXiv:2010.11125.

[3] Sneha Kudugunta, Isaac Caswell, Biao Zhang, Xavier Garcia, Derrick Xin, Aditya Kusupati, Romi Stella, Ankur Bapna, and Orhan Firat. 2023. Madlad-400: A multilingual and document-level large audited dataset. In Advances in Neural Information Processing Systems, volume 36, pages 67284--67296.

[4] Devendra Sachan and Graham Neubig. 2018. Parameter sharing methods for multilingual self-attentional translation models. In Proceedings of the Third Conference on Machine Translation: Research Papers, Belgium, Brussels.

[5] Firefox bilingual translation models. https://github.com/mozilla/firefox-translations-models

[6] Marcin Junczys-Dowmunt, Roman Grundkiewicz, Tomasz Dwojak, Hieu Hoang, Kenneth Heafield, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, André F. T. Martins, and Alexandra Birch. 2018. Marian: Fast neural machine translation in C++. In Proceedings of ACL 2018, System Demonstrations, pages 116--121.

[7] BrowserMT's Bergamot framework. https://github.com/browsermt/bergamot-translator

[8] Young Jin Kim, Marcin Junczys-Dowmunt, Hany Hassan Awadalla, Alham Fikri Aji, Kenneth Heafield, Roman Grundkiewicz, and Nikolay Bogoychev. 2019. From research to production and back: Ludicrously fast neural machine translation. In Proceedings of the 3rd Workshop on Neural Generation and Translation, pages 280--288.

[9] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.

[10] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomáš Mikolov. 2017. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427--431.

[11] Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C. Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022. COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 578--585.

[12] Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186–191, Belgium, Brussels. Association for Computational Linguistics.

[13] SDK. Qvac documentation. https://www.npmjs.com/ package/@qvac/sdk.