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
Minerva

Minerva

A language model trained further on technical content to solve maths and science problems

Minerva is a large language model from Google, introduced in a paper submitted on 29 June 2022. It starts from general natural language pretraining and is then trained on technical material so it can answer quantitative reasoning questions in maths, science and engineering without external tools.

Share on WhatsApp

  • Introduced by Google in a paper posted 29 June 2022
  • Pretrained on general text, then further trained on technical content
  • Solves quantitative problems with no external tools
  • Answered nearly a third of 200+ undergraduate science problems correctly
  • Covered physics, biology, chemistry, economics and other sciences

What it is

Minerva is a large language model built by Google and presented in June 2022. It is meant for quantitative reasoning: the kind of work involved in solving mathematics, science and engineering problems at college level. The model is first pretrained on general natural language data and then further trained on technical content. The team reports that it reaches state-of-the-art results on technical benchmarks while using no external tools such as calculators or computer algebra systems. They also tested it on more than two hundred undergraduate-level problems across physics, biology, chemistry, economics and other sciences, and it answered close to a third of them correctly.

How it came about

The work was described in the paper "Solving Quantitative Reasoning Problems with Language Models", first posted to arXiv on 29 June 2022 and revised on 1 July 2022. Fourteen researchers are listed as authors, including Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer and Guy Gur-Ari, who is named as the contact for the submission. The paper appears in the computation and language category, with cross-listings in artificial intelligence and machine learning.

Inside the model

Minerva follows a two-stage training recipe. It begins with pretraining on broad natural language data, the usual way a general language model learns to predict text. It is then further trained on technical material so that it handles mathematical and scientific writing better. The public description does not give layer counts, parameter totals, context length, tokenizer details or training compute.

What it can do

The team evaluated Minerva on technical benchmarks and reported state-of-the-art performance among models that use no external tools. They also built a set of more than two hundred undergraduate-level problems requiring quantitative reasoning, drawn from physics, biology, chemistry, economics and other sciences. On that set the model answered nearly a third correctly. The paper frames the result as progress on a gap: strong language models had generally struggled with quantitative reasoning at this level.

How to use it

The work is published as an arXiv paper with a PDF, and the submission links to code, data and media services that index the paper. No downloadable weights, licence terms or hosted model access are described.

Why it matters

Minerva is presented as a step towards language models that can do quantitative reasoning in science and engineering without leaning on external tools. Its evaluation set of over two hundred undergraduate-level problems across several sciences gives later work a concrete target for this kind of ability.

Questions people ask about Minerva

What is Minerva for?

It is a language model aimed at quantitative reasoning: solving mathematics, science and engineering problems of the kind found at college level. Google researchers built it to close the gap where earlier language models struggled with this sort of work.

How was it trained?

It was first pretrained on general natural language data, then further trained on technical content. That second stage is what gives it stronger handling of mathematical and scientific text compared with a purely general model.

Does it use a calculator or maths engine?

No. The researchers report that Minerva reaches its results on technical benchmarks without any external tools attached, working from the text of the problem alone.

How well does it do on real science problems?

The team tested it on more than two hundred undergraduate-level problems across physics, biology, chemistry, economics and other sciences. It answered nearly a third of them correctly.

Can I download or run Minerva?

The work is published as an arXiv paper with a PDF. No downloadable weights, licence terms or hosted model access are described in the available material.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Minerva yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2206.14858. 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.