Introducing HeliosLM: open Traditional Chinese (Taiwan) chat, speech recognition and safety models

Published November 8, 2026
Author: Chien-Hsin Lin

Today we are releasing the first HeliosLM models on Hugging Face. HeliosLM focuses on Taiwan: Traditional Chinese and local usage, Taiwanese Mandarin and Taiwanese Hokkien speech, and the scams and misinformation common in Taiwan. The chat and safety models are based on Google Gemma 4; the speech models are based on OpenAI Whisper large-v3.

What is released

Collection: https://huggingface.co/collections/chienhsinlin/helioslm-6aefeae35e2e4ff623907d13

Scores

All scores are self-run with the settings listed on each model card. The base-model column is the original base model measured with the same harness and items.

Speech recognition (CER %, lower is better)

Content safety (F1, unsafe class)

Chat E4B

Chat E4B is close to its base Gemma 4 E4B-it on general benchmarks; its focus is Taiwanese usage, a consistent identity, and neutral answers on contested topics.

Other languages

Per-language scores are on the model card. Other languages are mostly inherited from the base model and average about 1 point lower; Yoruba, Hausa and other African languages are weak.

HeliosLM HDL system (Verilog)

The HeliosLM HDL system is our generate / open-tool repair / select pipeline on NVIDIA's open Nemotron 3 Ultra (weights unmodified, OpenMDW-1.1). Pipeline scores; no HDL weights are released.

Neutrality

We hope HeliosLM presents facts and the main views of each side on political and historical issues, without taking a side.

• Chat E4B: on 155 political questions, the stance rate is 0%, and 96.1% are judged neutral.
• Guard E4B: ordinary opinions from 7 political camps are not marked unsafe; slurs and incitement against any side are flagged.

Known limitations

• Taiwanese Hokkien recognition is still weak, and the models output only Han characters, not Tâi-lô romanization.
• Chat E4B still produces one-sided answers on some items (16.8%, the same as the base model); we are still improving this.
• The models may produce incorrect or outdated information; please verify anything important yourself.

License

The Gemma-series models follow the Gemma Terms of Use; the Whisper-series models follow Apache-2.0.

Website: https://helioslm.pages.dev/