Translation models that run on 64 GB
18 models with published weights that fit in 64 GB — a workstation. Counted at four-bit quantisation, weights only, leaving the machine about a third of its memory and a gigabyte for context: room for roughly 67 billion parameters. A mixture of experts is counted in full, because it is held in full even though only a few experts compute.
A workstation, or a well-specified Mac. Seventy-billion-parameter models fit at four-bit, which is the class where local output stops being obviously worse than the hosted models people pay for. Loading takes real time and the machine will be warm.
Models and services aimed at moving text between languages. A general chat model will translate too, and often well; a dedicated one earns its place with glossaries, formality control, document formats it keeps intact, and rates that assume volume. Check which direction was measured — most quality figures are for translating into English, and the other direction is harder.
- Hy-MT2-30B-A3BTencent Hunyuan30.0B≈19.5 GB at 4-bit1 also selling it hosted
- translategemma-27b-itGoogle28.8B≈18.7 GB at 4-bit
- Voxtral Small 24B 2507Mistral AI24.3B≈15.8 GB at 4-bit4 also selling it hosted
- translategemma-12b-itGoogle13.2B≈8.6 GB at 4-bit
- granite-speech-3.3-8bIBM Granite8.0B≈5.2 GB at 4-bit
- Hy-MT2-7BTencent Hunyuan7.0B≈4.5 GB at 4-bit1 also selling it hosted
- Seed-X-PPO-7BByteDance Seed7.0B≈4.5 GB at 4-bit
- translategemma-4b-itGoogle5.0B≈3.2 GB at 4-bit
- Voxtral-Mini-3B-2507Mistral AI3.0B≈2.0 GB at 4-bit2 also selling it hosted
- madlad400-3b-mtGoogle2.9B≈1.9 GB at 4-bit
- seamless-m4t-v2-largeAI at Meta2.3B≈1.5 GB at 4-bit
- Hy-MT2-1.8BTencent Hunyuan1.8B≈1.2 GB at 4-bit1 also selling it hosted
- Whisper Large v3OpenAI1.5B≈1.0 GB at 4-bit2 also selling it hosted
- whisper-largeOpenAI1.5B≈1.0 GB at 4-bit
- whisper-large-v2OpenAI1.5B≈1.0 GB at 4-bit
- canary-1b-v2NVIDIA1.0B≈0.6 GB at 4-bit
- whisper-smallOpenAI0.2B≈0.2 GB at 4-bit
- whisper-tinyOpenAI0.0B≈0.0 GB at 4-bit