AlexaTM 20B
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
- Amazon
- Organisation type
- Industry
- Country
- United States of America
- Published
- 2 August 2022
- Authors
- Saleh Soltan, Shankar Ananthakrishnan, Jack FitzGerald, Rahul Gupta, Wael Hamza, Haidar Khan, Charith Peris, Stephen Rawls, Andy Rosenbaum, Anna Rumshisky, Chandana Satya Prakash, Mukund Sridhar, Fabian Triefenbach, Apurv Verma, Gokhan Tur, Prem Natarajan
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Language modeling, Translation, Question answering
- Numerical format
- BF16
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
- 19.8B
- Training data
- 1,319,000,000,000 tokens
- Batch size
- 2,000,000
See Table 1 on p.3 of the paper
See Table 2 on p.3 of the paper. 119B Wikipedia tokens + 1.2T mC4 tokens = 1319000000000 tokens
"We trained AlexaTM 20B for 120 days on 128 A100 GPUs for the total of 500k updates with the accumulated batch size of 2 million tokens"
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 × 10²³ FLOP
- How it was established
- Hardware
Training throughput is reported as 154 TFLOP/s - see p.5 of the paper. "We relied on an internal and optimized version of DeepSpeed that we have since open-sourced (Chiu & Zheng, 2022) to obtain training throughput of up to 154 TFLOPS/GPU on 16 AWS p4d.24xlarge compute instances." Accelerator compute days are reported as 15,360 days - see Table 17 on p.18 of the paper.
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
- NVIDIA A100
- Chips used
- 128
- Chip-hours
- 368,640
- Wall-clock time
- 2,880 hours (120 days)
- Hardware utilisation
- HFU 49.4%
- Power draw
- 102.6 kW
- Compute cost
- $267,943
- Cloud vendor
- Amazon Web Services
See p.5 of the paper: "We trained AlexaTM 20B for 120 days on 128 A100 GPUs..."
Training throughput reported at 154 TFLOPs/GPU, vs 312 TFLOPs/GPU for the A100s they use HFU = 154e12 / 312e12 = 0.4935
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
- API access
- Training code
- Unreleased
https://aws.amazon.com/about-aws/whats-new/2022/11/alexatm-20b-model-available-sagemaker-jumpstart/?nc1=h_ls
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Likely above 10²³ FLOP
- Yes
- Why it is tracked
- SOTA improvement
- Record confidence
- Confident
- Citations
- 90
The Abstract reports SOTA improvement on multiple benchmarks. "provides SOTA performance on multilingual tasks such as XNLI, XCOPA, Paws-X, and XWinograd"
Sources
Where this record came from and when it was last checked.
- Reference
- AlexaTM 20B: Few-Shot Learning Using a Large-Scale Multilingual Seq2Seq Model
- Last updated
- 25 May 2026
What the numbers mean
Where it came from
AlexaTM 20B was published by Amazon, in United States of America, in August 2022. industry is the category the publisher falls under.
It works in Language, and is recorded as doing language modeling, Translation, Question answering.
Its weights were never published, so it can only be reached through its provider. No graphics card changes that.
How it was trained
Producing it required around 2 × 10²³ FLOP of arithmetic, on NVIDIA A100, which is a statement about the training budget rather than about inference.
Around 1,319,000,000,000 tokens went into training it.
Its inclusion criterion is sOTA improvement.
Answers
AlexaTM 20B — common questions
Who created AlexaTM 20B?
AlexaTM 20B was published by Amazon, based in United States of America, categorised as industry.
When was AlexaTM 20B released?
AlexaTM 20B was published in August 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.
What is AlexaTM 20B used for?
AlexaTM 20B works in Language, and is recorded as handling language modeling, Translation, Question answering. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.
How much compute was used to train AlexaTM 20B?
Around 2 × 10²³ FLOP, on NVIDIA A100. 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.
What GPU do I need to run AlexaTM 20B?
None. AlexaTM 20B 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.
Is AlexaTM 20B open source?
No. AlexaTM 20B has not had its weights published, so it exists only as a service controlled by its owner.
How many parameters does AlexaTM 20B have?
AlexaTM 20B has 19.8B parameters. See Table 1 on p.3 of the paper. 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.
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.