gemma-4-E2B-it-text-only-non-thinking

A variant of tuandunghcmut/gemma-4-E2B-it-text-only with a 3-layer non-thinking enforcement in the chat template. The model weights are identical — only the chat template is modified to guarantee the model never enters thinking/reasoning mode.

---

3-Layer Non-Thinking Enforcement

---

Deploy with vLLM ✅

vllm serve tuandunghcmut/gemma-4-E2B-it-text-only-non-thinking \
  --dtype bfloat16 \
  --max-model-len 32768

Query via OpenAI-compatible API:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
response = client.chat.completions.create(
    model="gemma-4-E2B-it-text-only-non-thinking",
    messages=[{"role": "user", "content": "What is the capital of Vietnam?"}],
)
print(response.choices[0].message.content)
# → Direct answer, no thinking preamble

---

Use with 🤗 Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "tuandunghcmut/gemma-4-E2B-it-text-only-non-thinking",
    dtype=torch.bfloat16,
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained(
    "tuandunghcmut/gemma-4-E2B-it-text-only-non-thinking"
)

messages = [{"role": "user", "content": "Explain transformer architecture briefly."}]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

---

Compared to E4B variant

---

Source model

• Base: tuandunghcmut/gemma-4-E2B-it-text-only
• Original: google/gemma-4-E2B-it

License

Inherits the Gemma license.