Swift

Closed weights Intel Labs 56.8K parameters August 2023

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
Intel Labs
Organisation type
Industry
Country
United States of America
Published
30 August 2023
Authors
Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, Davide Scaramuzza

What it does

The problem areas the model was built for. A model can carry several of each.

Domain
Robotics
Task
Helicopter driving
Approach
Reinforcement learning

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
56.8K

The control network is an MLP with input dimension 31, two hidden layers of size 128, and an output of dimension 4. (31+1)*128+(128+1)*128+(128+1)*4 = 21124 Gate detector is a 6 layer U-net with 8*(3^3*3+1) + 16*(3^2*8+1) + 16*(3^2*16+1) + 16*(5^2*16+1) + 16*(7^2*16+1) + 16*(7^2*16+1) = 35680 35680 + 21124 = 56804

Training data
120,000,000 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
5.3 × 10¹⁶ FLOP

Policies are trained for a total of 1 × 108 environment interactions, which takes 50 min on a workstation (i9 12900K, RTX 3090, 32 GB RAM DDR5). Fine-tuning is performed for 2 × 107 environment interactions. 35.58 TFLOPS * 50 min * 60 s/min * 0.50 utilization = 5.337*10^16 FLOP

How it was established
Hardware

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 GeForce RTX 3090
Chips used
1
Wall-clock time
1 hours

50 minutes (training details, page 8)

Power draw
382 W

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

"Pseudocode for Swift detailing the training process and algorithms can be found in the file ‘pseudocode.zip’ on Zenodo at https://doi.org/10.5281/zenodo.7955278. To safeguard against potential misuse, the full source code associated with this research will not be made publicly available."

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Why it is tracked
SOTA improvement

"Our work marks the first time, to our knowledge, that an autonomous mobile robot achieved world-champion-level performance in a real-world competitive sport."

Record confidence
Likely
Citations
101

Sources

Where this record came from and when it was last checked.

Reference
Champion-level drone racing using deep reinforcement learning
Last updated
28 November 2025

What the numbers mean

Where it came from

Swift was published by Intel Labs, in United States of America, in August 2023. It comes out of industry.

It works in Robotics, and is recorded as doing helicopter driving.

Its weights were never published, so it can only be reached through its provider. No graphics card changes that.

Training and provenance

Training it took roughly 5.3 × 10¹⁶ FLOP of computation, on NVIDIA GeForce RTX 3090 — a measure of what producing the model cost, not of how fast it answers.

It was trained on about 120,000,000 tokens of text.

It is tracked in the underlying dataset for one reason in particular: sOTA improvement.

Answers

Swift — common questions

01

What GPU do I need to run Swift?

None. Swift 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.

02

Is Swift open source?

No. Swift has not had its weights published, so it exists only as a service controlled by its owner.

03

How many parameters does Swift have?

Swift has 56.8K parameters. The control network is an MLP with input dimension 31, two hidden layers of size 128, and an output of dimension 4. (31+1)*128+(128+1)*128+(128+1)*4 = 21124 Gate detector is a 6 layer U-net with 8*(3^3*3+1) + 16*(3^2*8+1) + 16*(3^2*16+1) + 16*(5^2*16+1) + 16*(7^2*16+1) + 16*(7^2*16+1) = 35680 35680 + 21124 = 56804. 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.

04

Who created Swift?

Swift was published by Intel Labs, based in United States of America, categorised as industry.

05

When was Swift released?

Swift was published in August 2023. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

06

What is Swift used for?

Swift works in Robotics, and is recorded as handling helicopter driving. These are the areas it was designed around; they describe intent rather than a hard boundary.

07

How much compute was used to train Swift?

Around 5.3 × 10¹⁶ FLOP, on NVIDIA GeForce RTX 3090. 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.

Source

Original publication

Record last updated 28 November 2025

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.