Fine-tuned FunctionGemma 270M for on-device function-calling inside Brinq's SmartPanel manufacturing-assistant demo. Shipped on the Synaptics Astra SL2619 SoC (2×Cortex-A55 @ 2 GHz, 1 TOPS Torq/Coral NPU, 2 GB DDR4) at Embedded World 2026.
What this model does
Given a user utterance and a list of tool declarations, the model emits one or more <start_function_call>call:NAME{...}<end_function_call> blocks or a plain natural-language reply. It was trained specifically to hit sub-500 ms decode latency on the SL2619 without giving up tool-selection accuracy on the SmartPanel domain.
Scope. The fine-tune is specific to the SmartPanel tool schema (maintenance procedures, alarm acknowledgement, photo capture, knowledge lookup). It's published here as prior art / starting checkpoint for the related Coral Dev Board physical-AI demo at Google IO 2026, not as a general-purpose function-calling model.
Files
File
Format
Size
Recommended use
smartpanel-v15-q4_k_m.gguf
GGUF Q4_K_M
253 MB
Production. Runs via llama.cpp on 2 GB / 2-core ARM targets.
smartpanel-v15-f16.gguf
GGUF F16
543 MB
Canonical checkpoint for re-quantization or further fine-tuning.
smartpanel-v12-q4_k_m.gguf
GGUF Q4_K_M
253 MB
Mid-production milestone.
smartpanel-v8-q4_k_m.gguf
GGUF Q4_K_M
253 MB
Device deployment milestone (what our SL2619 test boards have shipped with since Feb).
smartpanel-v4-q4_k_m.gguf
GGUF Q4_K_M
253 MB
First version with correct call: output format. Benchmark reference.
First correct output format. 84.2% domain accuracy, 142 ms avg latency on local llama-cpp.
v8
2026-02-24
call:
Deployed to Ollama on SL2619 test boards.
v8-moveworks
2026-02-26
call:
Variant trained with additional Moveworks-flavored examples. Not included here.
v8-fixed
2026-02-27
call:
Tokenizer hotfix.
v9–v13
Feb 27 – Mar 1
call:
Data curation + prompt-template iterations.
v15
2026-03-03
call:
Current production.
(v14 was trained but rolled forward into v15 before quantization — no separate artifact exists.)
Prompt format
FunctionGemma's native format. The tokenizer ships the <start_function_call>, <end_function_call>, <start_function_declaration>, <end_function_declaration>, <start_function_response>, <end_function_response>, and <start_of_turn> / <end_of_turn> special tokens.
<start_of_turn>user
You are a model that can do function calling with the following functions
<start_function_declaration>
declaration:set_led_color{description:<escape>Set RGB LED color<escape>,parameters:{...}}
<end_function_declaration>
<start_function_declaration>
declaration:play_buzzer{description:<escape>Sound the buzzer<escape>,parameters:{...}}
<end_function_declaration>
Turn the lights red and beep
<end_of_turn>
<start_of_turn>model
<start_function_call>call:set_led_color{color:<escape>red<escape>}<end_function_call><start_function_call>call:play_buzzer{pattern:<escape>beep<escape>}<end_function_call>
<end_of_turn>
1from llama_cpp import Llama
23llm = Llama(4 model_path="smartpanel-v15-q4_k_m.gguf",5 n_ctx=1024,6 n_threads=2,7 verbose=False,8)910prompt ="""<start_of_turn>user
11You are a model that can do function calling with the following functions
12<start_function_declaration>
13declaration:acknowledge_alarm{description:<escape>Dismiss the current alarm<escape>,parameters:{properties:{},required:[],type:<escape>OBJECT<escape>}}
14<end_function_declaration>
1516Ack the alarm
17<end_of_turn>
18<start_of_turn>model
19"""2021out = llm(prompt, max_tokens=128, temperature=0.1, stop=["<end_of_turn>"])22print(out["choices"][0]["text"])
Ollama
bash
1# Download the gguf, then:2cat> Modelfile <<'EOF'
3FROM ./smartpanel-v15-q4_k_m.gguf
4PARAMETER temperature 0.1
5PARAMETER num_ctx 2048
6PARAMETER stop "<end_of_turn>"
7PARAMETER stop "<end_function_call>"
8PARAMETER stop "<eos>"
9EOF1011ollama create smartpanel -f Modelfile
12ollama run smartpanel "Ack the alarm"
Benchmark (v3 / pre-v15, Jan 2026)
On SmartPanel domain (llama-cpp-python, Q4_K_M, local dev machine):
Model
Domain
Accuracy
Avg Latency
Output Format
Mobile Actions base
mobile
100 %
178 ms
call:
SmartPanel v1
smartpanel
66.7 %
355 ms
❌ declaration:
SmartPanel v2
smartpanel
36.8 %
135 ms
❌ partial output
SmartPanel v3 (precursor to v4)
smartpanel
84.2 %
142 ms
✅ call:
Mobile Actions (cross-domain)
smartpanel
66.7 %
159 ms
call:
v15 numbers forthcoming — benchmarks live in the Brinq internal repo.
Training
Base: unsloth/functiongemma-270m-it (BF16)
Method: LoRA fine-tune via Unsloth + TRL (SFTTrainer)
Hardware: A100 80GB (Docker, unsloth image)
Quantization: llama.cpp convert_hf_to_gguf.py --outtype f16 then llama-quantize ... 15 (Q4_K_M)
Training scripts, curated datasets, and eval harnesses live in Brinq's internal repo (not public). For the related Coral demo's dataset generators and fine-tune recipe (which are shipping public), see BrinqAI/coral-functiongemma-demo (currently private, planned public around Google IO 2026).
