Translation Models

SAVRN's Take

Most of our translation category is small, single-direction machinery. Of the 61 models, 35 come from the Helsinki-NLP Research Group, and their most-pulled entries are opus-mt pairs on the MarianMTModel architecture with a 512-token context: Dutch, French and Korean into English, English into Russian and German. AI at Meta adds 6, the T5 community 5 and Tencent 3. The download leader is t5-small at 24,966,975 pulls a month, nine and a half times t5-base at 2,626,991, then opus-mt-nl-en at 1,588,615 and nllb-200-distilled-600M at 1,135,320.

Running these comes down to memory, and not much of it. t5-small needs 0.1 GB at 16-bit, opus-mt-fr-en at 75M parameters needs 0.2 GB, t5-base 0.5 GB, so one accelerator holds dozens of pairs. The exception is Tencent's Hy-MT2-30B-A3B, a mixture-of-experts model with 30.1B parameters, a 262,144-token context and 33 languages in one checkpoint: 72.2 GB at 16-bit, 36.1 GB at 8-bit, 18.0 GB at 4-bit. Even so, the cheapest Index host for all four, the 30.1B included, is one MI300X at $1.85 an hour.

The license mix leans toward production: 40 of the 61 are Apache 2.0, 10 CC BY 4.0 and 4 MIT, against 4 CC BY-NC 4.0, 2 with none stated and 1 on Llama 3 terms. nllb-200-distilled-600M, fourth by downloads, sits under CC BY-NC 4.0, so keep it out of paid service, and opus-mt-en-de carries CC BY 4.0 while the other opus-mt leaders carry Apache 2.0, so read each pair's license instead of assuming the default. Only 15 of the 61 carry a parameter count; measure the rest at load time.

SAVRN Research, 2026-09-18