ACE-3-26B-A4B-Preview-260910

Model Description

ACE-3-26B-A4B-Preview-260910 is a preview release of APMIC's ACE-3 model family, built for Traditional Chinese (Taiwan) enterprise scenarios and agentic workflows.

The model is based on google/gemma-4-26B-A4B-it, a Mixture-of-Experts model with 26B total parameters and about 4B active parameters per token (the "A4B" in the name). APMIC has further optimized it to strengthen:

• Traditional Chinese output in Taiwan usage (terminology, phrasing and orthography)
• Both thinking and non-thinking response modes

Preview notice: this is a preview checkpoint intended for evaluation and feedback. It has not been through a full production release process. Please evaluate it on your own workloads before deploying.

---

Model Details

• Developed by: APMIC
• Model type: Gemma4ForConditionalGeneration (Transformers), sparse MoE (26B total / ~4B active)
• Base model: google/gemma-4-26B-A4B-it
• Language(s) (NLP): Traditional Chinese & English
• Weights precision: bfloat16 (not quantized)
• License: gemma (Google usage license; gated on Hugging Face)

---

Usage

messages = [{"role": "user", "content": "請用繁體中文簡單說明什麼是混合專家模型(MoE)。"}]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,   # True to let the model produce a reasoning block first
)

Transformers

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "APMIC/ACE-3-26B-A4B-Preview-260910"  # replace with the actual repo id

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [{"role": "user", "content": "請用繁體中文簡單說明什麼是混合專家模型(MoE)。"}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, enable_thinking=False,
    tokenize=True, return_dict=True, return_tensors="pt",
).to(model.device)

out = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Recommended sampling

The bundled generation_config.json uses temperature=1.0, top_p=0.95, top_k=64. These follow the base model's recommended settings.

Serving

The model can be served with any inference engine that supports Gemma 4 (for example vLLM). Because the weights are bfloat16, plan for roughly 52 GB of GPU memory for the weights alone, plus KV cache.

---

Intended Use and Limitations

Intended for

• Traditional Chinese (Taiwan) assistants and enterprise applications
• Function-calling and agent frameworks that need a locally deployable model
• Research and evaluation of the ACE-3 model family

Limitations

• This is a preview release; behavior may change in later ACE-3 versions.
• Like all LLMs, the model can produce incorrect or fabricated content. For high-risk use (financial, legal, medical), keep a human review or enterprise control layer in place.
• Tool-call outputs should be validated (schema and argument checks) before being executed.
• Performance on languages other than Traditional Chinese and English has not been specifically optimized.

---

License

This model is a derivative of Gemma 4 and is distributed under the Gemma Terms of Use. By using it you agree to Google's Gemma Terms of Use and Prohibited Use Policy.