羲和(HeliosLM)是一組以台灣繁體中文為主的開放模型。對話與安全模型以 Google Gemma 4 為基礎,語音辨識以 OpenAI Whisper large-v3 為基礎。
HeliosLM is a family of open models focused on Traditional Chinese (Taiwan). The chat and safety models are based on Google Gemma 4; the speech recognition models are based on OpenAI Whisper large-v3.
| Model | Base | What it does |
|---|---|---|
| helioslm-chat-e4b | Gemma 4 E4B-it | 對話(台灣繁體中文、英文) Chat |
| helioslm-guard-e4b | Gemma 4 E4B-it | 內容安全分類 Content-safety classifier |
| helioslm-chat-26b-a4b-GGUF | Gemma 4 26B-A4B-it | 對話 26B(GGUF,8 GB 顯卡可跑) Chat 26B, GGUF, runs on 8 GB GPUs |
| helioslm-hdl-e4b | Gemma 4 E4B-it | Verilog / SystemVerilog 硬體設計 HDL / RTL assistant |
| helioslm-hdl-e4b-GGUF | Gemma 4 E4B-it | HDL 模型 GGUF 版(8 GB 顯卡) HDL model, GGUF for 8 GB GPUs |
| helioslm-asr-tw-v5 | Whisper large-v3 | 語音辨識,以華語為主 ASR, mostly Mandarin |
| helioslm-asr-tw-v6m | Whisper large-v3 | 語音辨識,加強台語 ASR, stronger Taiwanese Hokkien |
| helioslm-tw-eval | — | 評測設定與分數 Evaluation settings and scores |
每個模型頁都列出訓練資料、評測設定與分數。所有分數都是我們自行測得。HDL 系統的分數列在本頁下方。 Each model card lists its training data, evaluation settings and scores. All scores are self-run. HDL system scores are below.
網站 Website: https://helioslm.pages.dev/
羲和 HDL 系統是我們的 Verilog / SystemVerilog 產生流程:產生答案、用開源工具檢查並修正、再挑出一個答案。底層模型是 NVIDIA 公開的 Nemotron 3 Ultra(550B,NVFP4,權重未修改),依其 OpenMDW-1.1 授權使用。這是流程的分數,不是 helioslm-hdl-e4b 模型的分數。
The HeliosLM HDL system is our Verilog / SystemVerilog generation pipeline: generate, check and repair with open-source tools, then select one answer. The base model is NVIDIA's open Nemotron 3 Ultra (550B, NVFP4, weights unmodified), used under its OpenMDW-1.1 license. These are pipeline scores, not scores of the helioslm-hdl-e4b model.
| Benchmark | HeliosLM HDL system |
|---|---|
| CVDP v1.1.0 non-agentic code generation (no_commercial, 302 problems) | 51.3% (155/302) |
| VerilogEval v2 | 0.859 |
| RTLLM v2.1 | 0.844 |
UVM 測試平台自動產生仍在開發中。UVM testbench generation is still in development.
所有 HDL 分數都是我們自己跑的,只用開源模擬器,尚未經第三方驗證。All HDL numbers are our own runs with open-source simulators only, not third-party verified. 網站 Website: https://helioslm.pages.dev/hdl/
地端單卡版(開放模型 gpt-oss-120b 加上模擬器回饋自動修正,一張 RTX PRO 6000)的實際輸出,每題都用開源模擬器和我們自己寫的 testbench 驗證:Verilog 同步 FIFO、SystemVerilog round-robin 仲裁器+SVA、在 OpenTitan UART 傳送模組裡找出並修正植入的 bug。題目、答案、testbench 與紀錄:chienhsinlin/helioslm-hdl-demo;網頁版:https://helioslm.pages.dev/hdl/demo.html ,HF 上也有同樣的頁面:HeliosLM HDL demo Space。UVM testbench 自動產生仍在開發中。
Unedited, simulator-verified outputs of the single-GPU on-prem HDL assistant: chienhsinlin/helioslm-hdl-demo · demo Space.
This repo also holds the original HeliosLM research code checkpoints: small, CPU-runnable reference implementations of common LLM building blocks (latent attention, mixture-of-experts routing, linear attention, speculative decoding) with oracle tests, plus a calibrated decision-layer experiment and its working paper (docs/). These toy checkpoints (8.5M to 360M parameters) are for learning and research, not for production use. Code: tonythetiger168/helioslm. A CPU demo runs in the HeliosLM Space.
This repo: Apache-2.0. Each model follows the license on its own card.