Minerva (540B)
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- Organisation type
- Industry
- Country
- United States of America
- Published
- 29 June 2022
- Authors
- Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, Vedant Misra
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Quantitative reasoning, Mathematical reasoning, Language modeling/generation, Question answering
- Approach
- Self-supervised learning
- Base model
- PaLM (540B)
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 540.4B
- Training data
- 26,000,000,000 tokens
"To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM)." Our approach is to start with the PaLM pretrained decoder-only transformer language models Chowdhery et al. (2022), and further train (finetune) them on our mathematical dataset using an autoregressive objective. Table 2 contains the main model and training hyperparameters. See Table 2
"Our models were trained on a dataset of 38.5B tokens" + PaLM upd 38.5B tokens - sie of the dataset, the model saw 26B tokens in 399k steps (see Table 2)
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 2.7 × 10²⁴ FLOP
- How it was established
- Hardware
- Fine-tuning compute
- 2.1 × 10²³ FLOP
Minerva was fine-tuned from PaLM using the same hardware. Assume the same model FLOPs utilization rate for pre-training and fine-tuning. "the 540B model was trained for 29 days on a v4-1024" PaLM pretraining time: 6144 TPU for 1200 hours + 3072 TPU for 336 hours = @8404992 TPU-hours Minerva finetuning time: 1024 TPU for 696 hours = 712704 TPU-hours So fine-tuning added 8.5% more compute. Minerva total compute = PaLM pretraining compute * (712704+8404992)/(8404992) = 2.7415*10^24 FLOP https://w…
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- Google TPU v4
- Chips used
- 1,024
- Chip-hours
- 712,704
- Wall-clock time
- 696 hours (29 days)
- Power draw
- 698.4 kW
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Closed — provider access only
- Model access
- Unreleased
- Training code
- Unreleased
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Frontier model
- Yes
- Foundation model
- Yes
- Likely above 10²³ FLOP
- Yes
- Why it is tracked
- SOTA improvement
- Record confidence
- Confident
- Citations
- 1,673
"We achieve state-of-the-art performance on MATH Hendrycks et al. (2021), GSM8k Cobbe et al. (2021), and a STEM subset of the MMLU Hendrycks et al. (2020) dataset"
Sources
Where this record came from and when it was last checked.
- Reference
- Solving Quantitative Reasoning Problems with Language Models
- Last updated
- 25 May 2026
What the numbers mean
Where it came from
Minerva (540B) was published by Google, in the country recorded as United States of America, during June 2022. It comes out of an organisation categorised as industry.
It works in the domain of Language, and is recorded as performing the task of quantitative reasoning, Mathematical reasoning, Language modeling/generation, Question answering.
Rather than being trained from scratch, it is derived from PaLM (540B). That is why it shares the base model's general shape and size.
Its weights were never published, so it can only be reached through its provider. No graphics card changes that.
Training and provenance
The training run consumed about 2.7 × 10²⁴ FLOP, on hardware recorded as Google TPU v4. That figure measures what producing the model cost, and has no bearing on how fast it answers.
Training consumed a corpus of around 26,000,000,000 tokens of text.
It is tracked in the underlying dataset for one reason in particular: sOTA improvement.
Answers
Minerva (540B) — common questions
Minerva (540B)— how much compute was used to train it?
Training consumed around 2.7 × 10²⁴ FLOP, on hardware recorded as Google TPU v4. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
Minerva (540B)— what GPU do I need to run it?
None. This is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
Minerva (540B)— is it open source?
No. Its weights have not been published, so it exists only as a service controlled by its owner.
Minerva (540B)— how many parameters does it have?
It has a parameter count of 540.4B. "To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM)." Our approach is to start with the PaLM pretrained decoder-only transformer language models Chowdhery et al. (2022), and further train (finetune) them on our mathematical dataset using an autoregressive objective. Table 2 contains the main model and training hyperparameters. See Table 2. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
Minerva (540B)— who created it?
It was published by Google, based in United States of America, an organisation categorised as industry.
Minerva (540B)— when was it released?
It was published in June 2022. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
Minerva (540B)— what is it used for?
It works in the domain of Language, and is recorded as handling the task of quantitative reasoning, Mathematical reasoning, Language modeling/generation, Question answering. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.