News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnToolformer

Toolformer

A language model that learns on its own when to call external tools through APIs.

Toolformer is a research language model that trains itself to call outside tools through simple APIs, deciding when a call helps and folding the answer back into its text predictions. It was introduced in a February 2023 paper by researchers at Meta AI and Sorbonne Université, and it showed stronger zero-shot results on many tasks while keeping its ordinary language skills.

Share on WhatsApp

  • Introduced in arXiv preprint 2302.04761, dated 9 February 2023
  • Learns tool calls with only a handful of demonstrations per API
  • Works with a calculator, Q&A system, two search engines, translation and a calendar
  • Improves zero-shot results without hurting core language modelling
  • Authors span Meta AI and Sorbonne Université

What it is

Toolformer is a language model altered so it can reach for outside programs while it writes. Presented in February 2023, it comes from a team working across Meta AI and Sorbonne Université. The system learns four things about each tool: whether calling it is worthwhile, the right moment to do so, which arguments to send, and how to weave the returned answer into the words that follow. Training needs only a small set of example calls per API. The paper reports better zero-shot results across a range of tasks, sometimes close to far bigger systems, with no loss of the model's basic text ability.

How it came about

The work appeared as arXiv preprint 2302.04761 on 9 February 2023, in the Computation and Language category. Eight researchers are listed on the paper, spanning Meta AI and Sorbonne Université. The abstract frames the project as an answer to a paradox: large models handle new tasks from a few examples yet fail at arithmetic and factual lookup, where small specialised tools do well. No earlier or later versions of the model are described.

Inside the model

The approach is self-supervised. For each API, the model is given a handful of demonstrations and then works out for itself which calls are useful. It learns to pick the tool, choose the moment, supply arguments, and blend the result into the next tokens it predicts. The tool set used in the work covers a calculator, a question-answering system, two separate search engines, a translation service and a calendar. Layer counts, parameter totals, training tokens, tokenizer details and compute are not given in the available description.

What it can do

The paper reports substantially better zero-shot performance on a variety of downstream tasks after the tool training. On several of these, results are described as competitive with much larger models. The authors also state that the model keeps its core language modelling ability, so ordinary text prediction is not degraded by learning to call tools. Concrete benchmark names and scores are not given in the available description.

How to use it

The work is presented as a research paper on arXiv, with a PDF and DOI. No downloadable weights, licence terms, hosted app or public API are described.

Why it matters

Toolformer set out a recipe for giving language models hands-on access to calculators, search and other services without hand-writing the calls. Its self-supervised scheme, needing only a few demonstrations per API, offered a template for later work on tool use and agents.

Questions people ask about Toolformer

What problem does Toolformer address?

Large language models pick up new tasks from a few examples but stumble on arithmetic and factual lookup. Toolformer lets the model call outside tools through APIs so it can lean on them where its own text prediction falls short.

How does it learn to use tools?

Training is self-supervised. A handful of demonstrations for each API is enough for the model to work out which calls help, when to make them, what arguments to pass, and how to fold the returned result into the words it generates next.

Which tools does it use?

The paper covers a calculator, a question-answering system, two different search engines, a translation system and a calendar.

Does using tools weaken its writing?

No, according to the authors. They report that the model keeps its core language modelling abilities alongside the improved zero-shot performance on downstream tasks.

Can I download or run Toolformer?

The available description covers only the arXiv paper and its PDF. It gives no weights, licence, app or API for the model.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Toolformer yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2302.04761. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.