Links & Datasets

Everything worth bookmarking about function calling. Links checked 2026-11-01. For benchmarks, see Benchmarks. To try models live, go to toolcalling.ai.

Official docs

Provider	Link
OpenAI	Function calling
Anthropic	Tool use
Google Gemini	Function calling · generateContent
Mistral	Function calling
Cohere	Tool use
AWS Bedrock	Tool use
Ollama	Tool calling
vLLM	Tool calling
Hugging Face	Chat templates & tools
MCP	Specification · SDKs · Announcement

More providers on Providers Compared.

Training datasets

Dataset	Content	License
xlam-function-calling-60k (APIGen)	60k verified query/tools/answer triples, 3,673 APIs	CC-BY-4.0, gated
APIGen-MT-5k	5k multi-turn agent trajectories	CC-BY-NC-4.0
glaive-function-calling-v2	~113k synthetic single- and multi-turn chats	Apache-2.0
hermes-function-calling-v1	Hermes <tool_call> format + JSON mode	Apache-2.0
hermes_reasoning_tool_use	Tool use with reasoning traces	Apache-2.0
ToolACE	~11k dialogues over 26k synthesized APIs	Apache-2.0
Toucan-1.5M	1.5M trajectories from real MCP servers	Apache-2.0
When2Call	Call / ask / refuse decisions	CC-BY-4.0
xlam-irrelevance-7.5k	No-call (irrelevance) examples	CC-BY-4.0
argilla/apigen-function-calling	xLAM 60k + Synth-APIGen merged for SFT	CC-BY-4.0
Nemotron-Agentic-v1 · Nemotron-Post-Training-v1	NVIDIA agentic and tool-calling SFT data	CC-BY-4.0
Agent-FLAN	Agent tuning data incl. negative samples	Apache-2.0
APIBench	Original Gorilla API-call data	Apache-2.0
ToolBench	16k+ RapidAPI APIs, ~126k solution paths (download via repo)	see repo

Always check the license before commercial use.

Open function-calling models

Model	Notes
Qwen3 · Qwen3.5 · Qwen3.6 · Qwen3-Coder	Apache-2.0, native Hermes-style tools (guide)
Llama 3.1 · 3.3 · 4 Scout	Llama licenses; prompt format
Mistral 7B v0.3 · Small 3.2 · Ministral 3 · Large 3	Native function calling
gpt-oss-20b / 120b	Apache-2.0, harmony format with built-in tools
xLAM-2 (1B-70B)	Multi-turn function calling; CC-BY-NC-4.0
Hammer 2.1	On-device function calling; CC-BY-NC-4.0
ToolACE-2.5-8B	Trained on ToolACE data
FunctionGemma 270M	Tiny on-device function caller
Hermes 3 · Hermes 4	Hermes format (repo)
Kimi K2 · GLM-4.5	Large agentic / tool-use models
Granite 3.3	Apache-2.0, native tools
Functionary · NexusRaven-V2 · Gorilla OpenFunctions v2 · watt-tool-70B · Arch-Function	Earlier and specialized function callers

Fine-tuning tooling

Tool	Tool-calling data docs
TRL	SFTTrainer · dataset formats (tools)
Axolotl	Conversation format with tools
LLaMA-Factory	Data format (function_call / observation)
Unsloth	Tool-calling guide · datasets guide
torchtune	Messages (no longer actively maintained)
OpenAI (hosted)	Supervised fine-tuning · Cookbook: fine-tuning for function calling
Google Vertex AI (hosted)	Supervised tuning · prepare data (functionCall)
Amazon Bedrock (hosted)	Custom models. Nova 2 SFT does not support tool sections; workaround
Together AI (hosted)	Data preparation
Mistral (hosted, legacy)	Fine-tuning

Frameworks & libraries

Library	Link
OpenAI Agents SDK	Tools
Claude Agent SDK	Overview
LangChain	Tools
LlamaIndex	Tools
Vercel AI SDK	Tools and tool calling
Pydantic AI	Tools
instructor	Structured outputs via tools
Outlines · XGrammar · llguidance	Outlines · XGrammar · llguidance: constrained decoding

Key papers

Foundations

• ReAct
• Toolformer
• Gorilla
• ToolAlpaca
• ToolLLM
• Tool Learning survey

Data synthesis

• APIGen
• ToolACE
• xLAM
• Hammer
• APIGen-MT
• Toucan
• Kimi K2
• Granite FC
• Hermes 3
• Llama 3 (tool-use section)
• Qwen3

RL for tool use

• ToolRL
• Nemotron-Research-Tool-N1

Evaluation

• BFCL
• τ-bench
• τ²-bench
• API-Bank
• T-Eval
• TaskBench
• StableToolBench
• ToolSandbox
• NESTFUL
• ComplexFuncBench
• ACEBench
• When2Call
• MCP-Bench
• MCP-Universe
• MCPMark
• MCP-Atlas

Constrained decoding

• Outlines
• XGrammar