Hy-MT2 1.8B

1.8B parameter open-weight model. Dense 1.8B translation model from Tencent's Hy-MT2 line (HunYuanDenseV1ForCausalLM, 32 layers, grouped-query attention with 4 KV heads, 262,144-token context, config.json read 2026-10-16), trained to translate between 33 languages plus five Chinese dialect and minority-language pairs rather than to chat. Apache 2.0. Parameter count: Ollama and the vendor's GGUF report 1.8B; the Hugging Face API counts 2.04B for the safetensors checkpoint, and the difference is exactly the 120,818 x 2,048 tied embedding matrix counted once. No Ollama library tag exists (registry checked 2026-10-16); Tencent publishes Q4_K_M, Q6_K and Q8_0 GGUFs at tencent/Hy-MT2-1.8B-GGUF (checked 2026-10-16) for llama.cpp and LM Studio. Measured on this project's own hardware from an in-house Q4_K_M conversion: 1.3 GB resident at 100% GPU on an RTX 3080, 259 to 282 tok/s across four runs (2026-09-05, docs/lab/radar-2026-09-05-model-validation.md) and 223 to 233 tok/s on 2026-10-16 while another service held 3.9 GB of the card (docs/lab/radar-2026-10-09-model-validation.md). Translation was spot-checked, not evaluated: single sentences into French, Japanese, Spanish and German came back accurate and idiomatic.