Gemma 4 E4B — Tool-Calling v0.2

A production-grade tool-calling model built on Google's Gemma 4 E4B-it. Fine-tuned on 78K curated examples to reliably call functions, select the right tool, and know when not to call.

4B parameters · 95% multi-tool accuracy · Runs on a single GPU · Apache 2.0

This model retains all general capabilities of Gemma 4 E4B-it — conversation, reasoning, knowledge. Tool-calling was added as an additional skill through fine-tuning.

---

Benchmarks

👁 Comparison with other models

👁 BFCL Results Breakdown

BFCL Leaderboard (Official)

BFCL Submission PR →

v0.1 → v0.2 Improvement

👁 v0.1 vs v0.2

---

Quick Start

Ollama (one command)

ollama run hf.co/roshangrewal/gemma4-e4b-toolcall-v02-gguf

vLLM (OpenAI-compatible API)

pip install vllm

vllm serve roshangrewal/gemma4-e4b-toolcall-v02 \
 --enable-auto-tool-choice \
 --tool-call-parser gemma4 \
 --max-model-len 8192 \
 --dtype float16

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

response = client.chat.completions.create(
 model="roshangrewal/gemma4-e4b-toolcall-v02",
 messages=[{"role": "user", "content": "What's the weather in Mumbai?"}],
 tools=[{"type": "function", "function": {
 "name": "get_weather",
 "description": "Get weather for a city",
 "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
 }}]
)

print(response.choices[0].message.tool_calls)

HuggingFace Transformers

from transformers import AutoProcessor, AutoModelForMultimodalLM
import torch

model = AutoModelForMultimodalLM.from_pretrained("roshangrewal/gemma4-e4b-toolcall-v02", torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained("roshangrewal/gemma4-e4b-toolcall-v02")

tools = [{"type": "function", "function": {
 "name": "get_weather",
 "description": "Get current weather for a city",
 "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
}}]

messages = [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What's the weather in Mumbai?"}]
text = processor.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))

LM Studio

1. Open LM Studio
2. Search roshangrewal/gemma4-e4b-toolcall-v02-gguf
3. Download and chat

---

Examples

---

Available Formats

---

Training

Data Sources

All data standardized to OpenAI format (messages + tools). 80% multi-turn conversations (avg 11.3 messages per example).

---

Limitations

• Outputs in Gemma 4 native format (vLLM auto-converts to JSON)
• Best accuracy on sequences up to 4K tokens (trained length); still works up to 32K
• 87.5% accuracy on knowing when NOT to call (may occasionally over-trigger)
• Not tested on non-English queries
• 4B model — less capable than 70B+ on complex multi-step reasoning

---

Citation

@misc{grewal2026gemma4toolcall,
 title={Gemma 4 E4B Tool-Calling v0.2},
 author={Roshan Grewal},
 year={2026},
 publisher={HuggingFace},
 url={https://huggingface.co/roshangrewal/gemma4-e4b-toolcall-v02}
}

Base model

google/gemma-4-E4B

google/gemma-4-E4B-it

(221)