MiniCPM5-1B-Agentic-Tooluse-v3-GGUF — Local Function-Calling LLM (llama.cpp / Ollama / LM Studio)
MiniCPM5-1B-Agentic-Tooluse-v3 is a
1-billion-parameter open-weight function-calling model you can run entirely offline on a CPU — no GPU, no cloud API, no data leaving your machine. It is quantized to GGUF format and works out of the box with
llama.cpp,
Ollama,
LM Studio, koboldcpp, and text-generation-webui.
If you are looking for a local LLM for tool calling, a small function-calling model for Raspberry Pi or a laptop, a private offline AI agent backbone, or a free alternative to GPT-4o / Claude function calling that runs on your own hardware, this is it.
74.67% exact-argument accuracy on a held-out 300-example benchmark — trained with QLoRA supervised fine-tuning followed by GRPO reinforcement learning, rewarding exact function-name and argument-value correctness. No GPU required at Q4_K_M.
Why this model
MiniCPM5-1B-Agentic-Tooluse-v3 is fine-tuned specifically to parse a tool schema and a natural-language user request, then emit a structured, correctly-named, correctly-valued function call — the exact skill that powers LangChain agents, LlamaIndex pipelines, AutoGen, CrewAI, MCP tool servers, ReAct loops, and home-automation assistants.
Unlike most small open tool-calling models that stop at supervised fine-tuning, this model goes further with GRPO reinforcement learning on top of the SFT checkpoint, specifically rewarding the two hardest parts of tool calling: choosing the right function name and getting every argument value exactly right.
Compared to GPT-4o / Claude for function calling: this model is 100% free, runs locally, keeps all data private, has zero per-call cost, and is fine-tunable — it trades some absolute accuracy for massive gains in cost, latency, and privacy.
Why this model
MiniCPM5-1B-Agentic-Tooluse-v3 is a compact 1B-parameter model fine-tuned specifically for agentic tool/function calling: it parses a tool schema plus a user request and reliably emits a structured, correctly-named, correctly-valued function call — the core capability behind LangChain agents, MCP servers, ReAct loops, home-automation assistants, and any app that needs an LLM to reliably drive external APIs and tools.
Unlike most small open tool-calling models, this one went through a two-stage pipeline: QLoRA supervised fine-tuning followed by GRPO reinforcement learning, specifically rewarding exact function-name and exact argument-value correctness.
Results
Evaluated on a held-out 300-example test slice drawn from a seeded shuffle of ToolACE (see Split integrity).
The base-model column is the same model with the same prompt and no adapter.
The published weights are SFT + GRPO (see GRPO / RLVR). The SFT column is kept because every
negative result below is measured against it.
| metric | v2 (previous release) | SFT retrain (pre-GRPO) | v3 = SFT + GRPO (published) |
|---|
parseable — output is a well-formed call | 0.9933 | 1.0000 | 1.0000 |
valid_name — name exists among the offered tools | 0.9700 | 0.9867 | 0.9867 |
expected_name — name matches gold | 0.9067 | 0.9567 | 0.9533 |
args_exact — every argument value matches gold | 0.6133 | 0.7367 | 0.7467 |
arg_key_overlap — F1 over argument keys | 0.8757 | 0.9422 | 0.9388 |
| mean of 5 | 0.8718 | 0.9245 | 0.9251 |
GRPO buys +0.0100 on args_exact, the metric that matters here, and gives back 0.0034 (one test example
each) on expected_name and arg_key_overlap. That trade is reported rather than hidden: the mean moves
only +0.0006, so this is a targeted gain on the hardest metric, not a broad improvement.
Full 8-metric benchmark (held-out test set, n=300)
This table mirrors the evaluation format from v2 and shows Base, v2, and v3 side-by-side
across all 8 metrics using a single consistent harness and held-out test slice:
| Metric | Base MiniCPM5-1B | v2 (previous release) | v3 (this model) | Delta (v2 → v3) |
|---|
| parseable_rate | 0.0133 | 0.9933 | 1.0000 | +0.0067 |
| valid_name_rate | 0.0133 | 0.9700 | 0.9867 | +0.0167 |
| expected_name_rate | 0.0133 | 0.9267 | 0.9533 | +0.0267 |
| args_exact_rate | 0.1500 | 0.6533 | 0.7467 | +0.0934 |
| arg_key_overlap | 0.0033 | 0.7517 | 0.9388 | +0.1871 |
| no_schema_copy_rate | 1.0000 | 1.0000 | 0.9967 | -0.0033 |
| no_repetition_rate | 0.9967 | 1.0000 | 0.3400 | -0.6600 |
| stopped_cleanly_rate | 0.0000 | 0.1500 | 0.0000 | -0.1500 |
What the additional metrics mean:
no_schema_copy_rate — the model did not copy the tool schema's own field description
verbatim into an argument value.no_repetition_rate — the completion did not contain a duplicated function-call block or
degenerate repeated-phrase loop. This model has a known weakness here: it often continues
generating filler content after the tool call completes. Use a parser that extracts the first
completed <function>...</function> block.stopped_cleanly_rate — the model naturally stopped immediately after the completed
</function> tag with no trailing tokens. Use a parser that treats the first completed
<function>...</function> block as the action boundary — do not rely on natural end-of-generation.
Available quantizations
| File | Quant | Size | Best for |
|---|
MiniCPM5-1B-Agentic-Tooluse-v3.F16.gguf | F16 | ~2.02 GB | Maximum quality, GPU or high-RAM CPU inference |
MiniCPM5-1B-Agentic-Tooluse-v3.Q8_0.gguf | Q8_0 | ~1.07 GB | Near-lossless quality, recommended default for most users |
MiniCPM5-1B-Agentic-Tooluse-v3.Q4_K_M.gguf | Q4_K_M | ~656 MB | Smallest, fastest — best for edge devices, phones, and CPU-only/low-RAM machines |
Quickstart
llama.cpp:
1
2./llama-cli -m MiniCPM5-1B-Agentic-Tooluse-v3.Q8_0.gguf -p "Your prompt with tool schema here"
3
llama-server (OpenAI-compatible API, works with most agent frameworks):
1
2./llama-server -m MiniCPM5-1B-Agentic-Tooluse-v3.Q4_K_M.gguf --port 8080
3
Ollama:
1
2# Create a Modelfile:
3
4# FROM ./MiniCPM5-1B-Agentic-Tooluse-v3.Q8_0.gguf
5
6ollama create minicpm5-tooluse-v3 -f Modelfile
7
8ollama run minicpm5-tooluse-v3
9
LM Studio: just download one of the .gguf files above directly through the LM Studio search/download UI.
Ideal use cases
-
Fully local / offline / private AI agents (no data leaves your machine)
-
Home automation and smart-home voice assistants
-
Mobile, browser-extension, and embedded/IoT tool-calling agents
-
Cost-sensitive, high-volume backend services that can't afford large-model API costs per call
-
Drop-in function-calling backbone for LangChain, LlamaIndex, AutoGen, CrewAI, and MCP-based agent stacks
-
Hobbyist and researcher experimentation with small-model agentic reasoning
FAQ
Which quant should I use? Q8_0 for the best quality-to-size tradeoff on most machines; Q4_K_M if you need the smallest possible footprint or are running on a phone/Raspberry Pi-class device; F16 if you have plenty of RAM/VRAM and want maximum fidelity.
Do I need a GPU? No — that's the point of this model. All three quantizations run well on CPU; a GPU just makes it faster.
How was this trained? QLoRA supervised fine-tuning on tool-calling trajectories, followed by GRPO (Group Relative Policy Optimization) reinforcement-learning refinement targeting exact argument correctness.
Base model architecture
MiniCPM5-1B uses a standard LlamaForCausalLM architecture:
| Property | Value |
|---|
| Parameters (total) | 1,080,632,832 |
| Parameters (non-embedding) | 679,552,512 |
| Architecture | LlamaForCausalLM |
| Layers | 24 |
| Attention heads (GQA) | 16 Q / 2 KV |
| Context length | 131,072 tokens |
| Training | SFT → RL (GRPO) fine-tune on openbmb/MiniCPM5-1B |
Thinking mode
MiniCPM5-1B has a built-in <think>...</think> chat template. The same checkpoint can act as a fast assistant or a deliberate chain-of-thought reasoner — controlled by a single flag:
1# Fast mode — recommended for tool calling (thinking OFF)
2prompt = tokenizer.apply_chat_template(
3 messages, tools=tools, add_generation_prompt=True,
4 enable_thinking=False,
5 tokenize=False,
6)
7
8# Reasoning mode (thinking ON — NOT recommended for tool calling)
9prompt = tokenizer.apply_chat_template(
10 messages, tools=tools, add_generation_prompt=True,
11 enable_thinking=True,
12 tokenize=False,
13)
Important: always use enable_thinking=False for tool/function calling. With thinking ON the model spends its token budget inside <think>...</think> and may not reach a completed function call. All benchmark numbers in this card use thinking OFF.
Citation
If you use this model, please cite the base model paper:
1@article{minicpm4,
2 title = {MiniCPM4: Ultra-Efficient LLMs on End Devices},
3 author = {MiniCPM Team},
4 journal = {arXiv preprint arXiv:2506.07900},
5 year = {2025}
6}
And the ToolACE dataset used for fine-tuning:
1@article{toolace,
2 title = {ToolACE: Winning the Points of LLM Function Calling},
3 author = {Liu, Ying and others},
4 journal = {arXiv preprint arXiv:2409.00920},
5 year = {2024}
6}
ModelScope
The base model is also available on ModelScope (for users in China and East Asia):
(The fine-tuned adapter/GGUF builds are currently HuggingFace-only.)
Related repos
v3 model family (this release)
Previous releases
Base model
Built on
MiniCPM5-1B by OpenBMB, fine-tuned for agentic tool/function calling and refined with GRPO reinforcement learning.
Limitations
