Instructions to use chienhsinlin/helioslm-chat-e4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use chienhsinlin/helioslm-chat-e4b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="chienhsinlin/helioslm-chat-e4b")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("chienhsinlin/helioslm-chat-e4b")
model = AutoModelForMultimodalLM.from_pretrained("chienhsinlin/helioslm-chat-e4b", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use chienhsinlin/helioslm-chat-e4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "chienhsinlin/helioslm-chat-e4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "chienhsinlin/helioslm-chat-e4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/chienhsinlin/helioslm-chat-e4b

SGLang

How to use chienhsinlin/helioslm-chat-e4b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "chienhsinlin/helioslm-chat-e4b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "chienhsinlin/helioslm-chat-e4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "chienhsinlin/helioslm-chat-e4b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "chienhsinlin/helioslm-chat-e4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use chienhsinlin/helioslm-chat-e4b with Docker Model Runner:
```
docker model run hf.co/chienhsinlin/helioslm-chat-e4b
```

羲和 HeliosLM Chat E4B

羲和（HeliosLM）Chat E4B 是以 Google Gemma 4 E4B-it 為基礎的對話模型。我們在 Gemma 4 E4B 上以台灣繁體中文、英文等公開資料做持續預訓練，再把這個差異以 chat vector 方式加回 Gemma 4 E4B-it（沒有額外 SFT），並在 chat template 內建一段羲和系統提示（自我介紹、台灣用語、爭議議題保持中立）。使用者自訂的 system prompt 會接在這段提示之後。

HeliosLM Chat E4B is a chat model based on Google Gemma 4 E4B-it. We continued pre-training Gemma 4 E4B on public Traditional Chinese (Taiwan), English and other data, added that difference to Gemma 4 E4B-it as a chat vector (no extra SFT), and built a default HeliosLM system prompt into the chat template.

參數量 Parameters: about 8.0B total (Gemma 4 E4B architecture)
授權 License: Gemma Terms of Use. Gemma is provided under and subject to the Gemma Terms of Use.
建議使用 vLLM 或 transformers；思考模式以 enable_thinking=True 開啟（chat template 參數）。

評測 Evaluation (2026-11-04, self-run)

All numbers below were measured by us with the same open settings: vLLM 0.33, bf16, thinking on, temperature 1.0, top_p 0.95, top_k 64. They are self-reported and not verified by the benchmark owners. Scores of other models are not listed here; see each benchmark's leaderboard.

Benchmark	Score	Setting
MMLU-Pro	69.51	all 12,032 test items, single sample, 0-shot CoT, max 16,384 new tokens
GSM8K	93.71	all 1,319 test items, single sample, 0-shot, benchmark eval.yaml prompt, max 8,192 new tokens
AIME 2026	45.83	30 problems, mean of 4 samples, max 32,768 new tokens, math-verify
HMMT Feb 2026	31.82	33 problems, mean of 4 samples, max 32,768 new tokens, math-verify
HumanEval	95.12	164 problems, pass@1, single sample
TMMLU+ (Twinkle Eval 2.10, pattern, temp 0, max 1,024 tokens, thinking off)	41.37	full set, 1 run; measured with the previous system prompt

For reference, Google reports MMLU-Pro 69.4 and AIME 2026 42.5 for Gemma 4 E4B-it on its own model card (different harness). Our model is close to Gemma 4 E4B-it on these general benchmarks; it is not a large improvement over its base. The Twinkle TMMLU+ run used a short 1,024-token limit and is lower than our own reasoning-mode TMMLU+ measurement; treat it as a lower bound.

身分與中立性檢查 Identity and neutrality probes

Measured on 200 identity prompts (says HeliosLM: 100%) and 155 political prompts (judge: Gemma 4 26B-A4B-it):

自稱羲和 Says it is HeliosLM: 100%
中立判定 Judged neutral on political topics: 96.1% (sovereignty 100%, 統獨 100%); Gemma 4 E4B-it on the same probes: 93.5%
表態率 Takes a side (unification / independence): 0% (0 of 155)
被判為帶有單方官方說法 Judged to use one-sided official framing: 16.8% (Gemma 4 E4B-it: 16.8%, Gemma 4 26B-A4B-it: 18.1% with the same judge); highest on questions about Chinese leadership and Taiwan's institutions. We are still working on this.
拒答 Refusals: 2.6%

模型可能產生錯誤或過時資訊，請自行查證重要內容。The model can make mistakes; verify important information.

Downloads last month: 33

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for chienhsinlin/helioslm-chat-e4b

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Finetuned

(461)

this model