FTW (For The Win)

Training consumed around 3.5 × 10¹⁹ FLOP. 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.

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.

No. Its weights have not been published, so it exists only as a service controlled by its owner.

It has a parameter count of 126M. Architecture described in figure S11 of the supplement The architecture includes modules for visual embedding, reward prediction, recurrent processing, policy, baseline and pixel control. Input is 84x84x3 pixels as seen in figure S10 of the supplement "We elected to use a resolution of 84x84 pixels as in previous related work in this environment. Each pixel is represented by a triple of three bytes" Visual embedding (84x84x3 -> 256) 32*(8*8*3+1)+64*(4*4*32+1)+64*(3*3*64+1)+64*(3*3*64+1) + (84/(S^4)*84/(S^4)*64+1)*256 Note there is no information about the stride S used in the convolutions; we assume S = 1 Reward prediction (256 -> 3) (256+1)*128 + (128+1)*3 Recurrent processing (n-> 512) VU1 (256 -> 512) 4*(799+2*32)*((512+(32*2) + 3*32 + 5*2 + 3)+(799+2*32)+1) + 2*(256+1)*256 VU2 (512 -> 512) 4*(512+2*32)*((512+(32*2) + 3*32 + 5*2 + 3)+(512+2*32)+1) + 2*(256+1)*256 LSTMs usually have 4*(n*m+n*n+n) parameters, where n=input size and m=output size. This DNS + LSTM takes as input the concatenation of the previous layer of size n and R read vectors of size W=32; and outputs m units plus an interface vector of size (W*R) + 3*W + 5*R + 3, for a total of about 4*(n+R*W)*((m+(W*R) + 3*W + 5*R + 3)+(n+R*32)+1) parameters I assume R=2 since that seems implied by the previous paper (?) The first VU has as input the visual embedding (size 256), the previous action (size 540) and the previous reward (size 3), for a total size of 256+540+3 = 799. The output is size 512. The second VU has input size 512 and output size 512 The DNC memory architecture is described in https://www.nature.com/articles/nature20101.epdf Policy (512 -> 5x3x3x3x2x2) 6*(512+1)*256 + (256+1)*5 + 3*(256+1)*3 + 2*(256+1)*2 Baseline (512+1)*256 + (256+1)*1 Pixel control (512+1)*32*7*7 + 32*(9*9+1) + 5*(4*4+1) + 3*2*(4*4+1) + 2*2*(4*4+1) + 1*(4*4+1) "we trained independent pixel control policies for each of the six action groups". 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.

It was published by DeepMind, based in United Kingdom of Great Britain and Northern Ireland, an organisation categorised as industry.

It was published in July 2018. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

It works in the domain of Games, and is recorded as handling the task of capture the flag. These are the areas it was designed around; they describe intent rather than a hard boundary.