Hy-MT1.5-1.8B-1.25bit: A 440MB Offline Phone Translation Model
Melvin Vivas · X video post · 2026-04-30 · Open on X
Topics: LLM Fundamentals, Industry Trends & Job Market · Level: intermediate
Summary
An open-sourced 1.8B-parameter translation model, quantized to 1.25 bits so it takes only 440MB, runs fully offline on a phone. It supports 33 languages and, according to its makers, beats Google Translate on standard benchmarks.
Key points
- 1.8B parameters, quantized to 1.25 bits, for a 440MB model.
- Runs fully offline on a phone.
- Supports 33 languages.
- Claimed to match commercial translation APIs and 235B-scale models on standard benchmarks.
- Shows how far extreme quantization can push small on-device models.
Resources mentioned
- Hy-MT1.5-1.8B-1.25bit · tool · huggingface.co · free
An open-source 1.25-bit translation model, 440MB, covering 33 languages for offline use on phones.
Try this
- Build an offline mobile translation app on top of Hy-MT1.5-1.8B-1.25bit.
More in LLM Fundamentals
- MiniCPM-V 4.6 1.3B: Small Open-Source Vision/OCR Model for Edge Devices
- Enable Gemma 4 MTP Speculative Decoding on iPhone
- Gemma 4 Gets Up to 3x Faster with MTP Drafters
- Run Qwen3.6-27B Locally in 18GB RAM with Unsloth GGUFs
- Qwen3.6-27B: Dense Open Model for Local Agentic Coding
- Gemma 4 Now Available Through the Gemini API (Dev Use Only)