Home › Language Models › watt-tool-70B

watt-tool-70B - AI Language Models Tool

Overview

watt-tool-70B is a 70–71B parameter instruction-following LLM fine-tuned from Meta LLaMa-3.x (listed as meta-llama/Llama-3.3-70B-Instruct) and optimized specifically for tool usage, function calling, and multi-turn agent workflows. The model card and README emphasize supervised fine-tuning plus “Direct Multi-Turn Preference Optimization” (DMPO) and chain-of-thought style data to improve multi-turn tool orchestration and function-selection decisions, making the model suitable for workflow builders and agent platforms such as Lupan and Coze. ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B)) Practically, watt-tool-70B is intended to produce structured function calls (tool selection + arguments), to withhold calls when none are appropriate, and to maintain conversational context across multiple turns. The project is published under an Apache-2.0 license and is distributed in BF16 weights on Hugging Face, with community-maintained quantized variants (GGUF / GPTQ) available for local/edge use. The authors and community posts also note that watt-tool-70B achieved top performance on the Berkeley Function-Calling Leaderboard (BFCL) at times, though leaderboard rankings have continued to evolve. ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B))

Model Statistics

  • Downloads: 95,770
  • Likes: 118
  • Parameters: 70.6B

License: apache-2.0

Model Details

Architecture & lineage: fine-tuned from meta-llama/Llama-3.3-70B-Instruct (LLaMa-3 family) into a 70B-class causal language model specialized for function-calling and tool use. The Hugging Face model card lists the model as ~71B parameters and BF16 tensor weights. ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B)) Training and optimization: the maintainers report Supervised Fine-Tuning (SFT) on curated multi-turn tool-use data, use of Chain-of-Thought (CoT) style synthesis, and Direct Multi-Turn Preference Optimization (DMPO) to improve multi-step agent decisions. The model card links to an associated paper describing the DMPO approach (arXiv:2406.14868). ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B)) Capabilities and intended use: optimized for accurate function selection, multi-step tool orchestration, and retaining context across dialog turns. Example target applications include automated workflow builders, orchestration layers that must call multiple APIs, multi-turn data-extraction pipelines, and research into agentic LLM behavior. The model outputs structured function-call text when provided with function definitions in prompts and is designed to avoid spurious calls when no tool is appropriate. ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B)) Formats & deployment: official weights are distributed via Hugging Face (safetensors, BF16). Community-maintained quantized builds (GGUF, GPTQ) exist for local inference (see several community repos for Q3/Q4/GGUF variants). The model card notes it is not currently deployed by an inference provider (so hosted API pricing is provider-dependent). ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B))

Key Features

  • Fine-tuned for accurate function/tool selection in multi-step workflows.
  • Optimized multi-turn dialogue: retains context across chained tool calls.
  • Structured function-call outputs compatible with JSON-like tool definitions.
  • Trained with SFT and Direct Multi-Turn Preference Optimization (DMPO).
  • Distributed in BF16 safetensors with community GGUF/GPTQ quantizations.
  • Apache-2.0 licensed for permissive reuse and commercial integration.

Example Usage

Example (python):

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "watt-ai/watt-tool-70B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# Example: provide system+user messages and a JSON-style list of functions in the prompt
system = "You are an assistant that returns function calls when appropriate. Only emit valid calls when necessary."
user = "Get me the latest weather for San Francisco and schedule a meeting tomorrow at 10am if storm expected."
functions = "[{\"name\": \"weather_forecast\", \"args\": {\"location\": \"San Francisco, CA\", \"days\": 1}}, {\"name\": \"calendar.create_event\", \"args\": {\"title\": \"Meeting about weather\", \"time\": \"2026-03-04T10:00:00\"}}]"

prompt = f"System: {system}\nUser: {user}\nFunctions: {functions}\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# Note: the Hugging Face model card includes usage examples and a chat-template example; adapt tokenization/generation settings for your runtime. ([huggingface.co](https://huggingface.co/watt-ai/watt-tool-70B))

Benchmarks

Berkeley Function-Calling Leaderboard (BFCL): Model card reports state-of-the-art/top performance on BFCL (leaderboard rankings have since evolved) (Source: https://huggingface.co/watt-ai/watt-tool-70B)

Hugging Face downloads (last month): 95,770 (downloads last month, listed on model page) (Source: https://huggingface.co/watt-ai/watt-tool-70B)

Hugging Face likes / community stars: 118 likes on Hugging Face (Source: https://huggingface.co/watt-ai/watt-tool-70B)

Model size (parameters): ≈71B parameters (model card) (Source: https://huggingface.co/watt-ai/watt-tool-70B)

Quantized community builds: Multiple GGUF/GPTQ quantizations available (Q3/Q4 variants) for local inference (Source: https://huggingface.co/bartowski/watt-ai_watt-tool-70B-GGUF)

Last Refreshed: 2026-03-03

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.