Our inference engine.
Kokonut Island runs FYSK O and FYSK C on your machine, without internet.
NVIDIA RTX Spark compatible
In-house, and it changes everything.
The AI runs on your machine, not in the cloud. For an accounting firm, that changes four things.
-
Nothing leaves the firm.
Your documents are read and booked on the premises. Nothing goes to an AI provider.
Confidentiality
-
A binder in a few minutes.
300 documents read in under a minute, booked in under six. Measured on a DGX Spark.
Speed
-
Works without internet.
A network outage stops nothing. Your documents keep being read and booked.
Autonomy
-
No cloud bill by volume.
No AI provider paid per document. Your hardware, bought once, does all the work.
Cost
Everything runs on the machine.
On Kokonut Station or on your own, with no outgoing connection at all.
- Neither documents nor answers in the logs
- Internet only to install and update
Three figures to remember.
Two throughputs, measured on a DGX Spark, and the graphics memory Kokonut Island needs.
- documents per minute
- 390
- FYSK O
- documents per minute
- 58
- FYSK C
- of graphics memory is enough
- 14 GB
- Kokonut Island
On the hardware you choose.
A single NVIDIA card is enough, from 14 GB of graphics memory: the one in a PC, a workstation or a server.
Your current card or your next one: the last three NVIDIA generations.
Blackwell
The current generation, from the DGX Spark to the data centre.
- RTX Spark
- DGX Spark
- GeForce RTX 50 Series
- RTX PRO Blackwell Series
- DGX Station GB300
- B200
- B300
Hopper
The data centre cards.
- H100
- H200
Ada
The previous generation, from PC to server.
- GeForce RTX 40 Series
- RTX Ada Series
- L40S
- L4
Soon on your Mac, with MLX.
Chips
- M1
- M2
- M3
- M4
- M5
Soon on AMD cards, from Radeon to the data centre.
CDNA 4
- Instinct MI350 Series
CDNA 3
- Instinct MI300 Series
RDNA 4
- Radeon RX 9000 Series
- Radeon AI PRO R9700
RDNA 3
- Radeon RX 7000 Series
- Radeon PRO W7000 Series
RDNA 3.5
- Ryzen AI Max Series
vLLM
The engine that serves the models: documents are processed in parallel, in batches.
island init
- Machine NVIDIA Blackwell
- Engine vLLM ready
- FYSK O loaded
- FYSK C loaded
- Network offline, no outgoing connection
- Island ready, waiting for Kokonut Wave
Kokonut Island takes care of everything.
Once installed, you never touch it again. Kokonut Wave calls it, Island does the rest.
-
Nothing to start
A document arrives? The models load, then free the memory.
-
Nothing to watch
A model stops? Kokonut Island restarts it by itself.
-
Nothing to schedule
Too many documents at once? Each one waits its turn.
-
Nothing to choose
Read or book? Each document goes to the right model, FYSK O or FYSK C.
Frequently asked questions
Do I have to buy Kokonut Station?
No. Kokonut Island comes with Kokonut Station. On your own machine, it is offered under licence: get in touch with us.
Which models does it run?
FYSK O, which reads the documents, and FYSK C, which does their accounting classification and the entry. It will also be compatible with the FYSK models currently in development.
With Kokonut Station, or on your machine.
Kokonut Island comes with Kokonut Station, turnkey, or under licence on your own hardware.
-
Kokonut Station
Turnkey: the machine, Kokonut Wave and the FYSK models, ready to use.
Kokonut Island
-
Your machine
An NVIDIA card from 14 GB of graphics memory, or an NVIDIA RTX Spark.
Kokonut Island
