TranslatePsy Open-Sources African-Language Translation Models, With Quantized 0.8B to 4B Versions for Offline Use

Tether AI Research has officially released TranslatePsy, a collection of machine translation resources for African languages. AfriSLM comes in approximately 0.8B, 2B, and 4B sizes. Each model was created through full-parameter supervised fine-tuning of Qwen3.5 and supports bidirectional translation between English and 19 sub-Saharan African languages while retaining a multi-turn conversation format. All three sizes are also available as Q8_0 and Q4_K_M GGUF files for deployment through llama.cpp, Ollama, or compatible tools. The models and code are licensed under Apache 2.0, while the public synthetic dataset uses CC BY-NC 4.0; anyone retraining the models or reusing the data must review the licenses separately.

The training mixture includes quality-filtered synthetic parallel corpora, multilingual instruction data, and parallel data covering 38 Asian and European languages to mitigate catastrophic forgetting. The publicly released synthetic portion contains approximately 215.6 million bidirectional examples and about 32.37 billion tokens before fine-tuning. The models were trained for one epoch on 32 H100 GPUs using DeepSpeed ZeRO-2, with sequences longer than 2,048 tokens excluded. The paper claims that unified quality-estimation filtering can remove up to 96% of tokens without reducing test quality. The project therefore focuses not only on shrinking models but also on making data selection for low-resource translation reproducible.

In bidirectional evaluations across 19 languages on FLORES-200, BOUQuET, and SMOL, the 4B model achieved macro-average SSA-COMET scores of 0.6143, 0.6391, and 0.5136, respectively. According to the authors’ paired bootstrap tests, the smaller models outperform larger baselines such as TranslateGemma-27B and Qwen3.5-122B-A10B in most settings. The 4B Q4_K_M model’s scores declined only slightly, to 0.6138, 0.6377, and 0.5121, suggesting that quantized deployments may preserve most performance measured by automated evaluation. A separate AfriNano track uses Marian/Bergamot to reduce model sizes to between 17M and 43M parameters. It supports eight African languages and is better suited to browsers, smartphones, and CPU-based edge devices.

The limitations are equally clear. Most training translations were synthesized by other models and may inherit errors from the NLLB teacher model and the quality metrics. The paper includes no human evaluation by native-language experts, while macro averages can obscure weaknesses in individual languages, dialects, and translation directions. Engineering teams should first build human-evaluated test sets covering their target regions, code-switching, specialized terminology, and long documents. Current scores alone are not sufficient for high-risk medical, legal, or immigration use cases.