All Services are restored on our new bare metal systems

services suspended for the next 12 hours

e now supports

- nemotron NVFP4
- GLM 5.3 NVFP4

for your GB300 station

We took time and care to make sure our NVFP4 was done right

throughput beat the hell out of everyone

your loss if you own a station and dont use actual.

I’m not sure which is more goated: our inference engine spitting out 10k tok/s or @jack15032’s graphic design

The team at @actualinc has had early access to Nemotron 3.5 Super VL and used it to show phenomenal performance of the model on DGX Station using actual. It’s been a privilege to work with them these past few months. Their blog is cross linked in the official launch blog 🚀💚

Actual Computer turns your NVIDIA DGX Station into a department-scale AI factory.

Read the full case study:
actual.inc/company/blog/c…

One DGX Station can serve a whole department, with each person running their own swarm.

Your whole team reaches the station from anywhere through api.actual.inc.

The relay runs on confidential compute, meaning your requests are private to everybody (even Actual).

Download Actual Computer today.

Use it to run Nemotron 3.5 Super VL on your GB300

actual.inc

a few notes on local AI

and our computers that run it

around fifteen months ago, i led the pre-seed round for a weird local AI project

it was called @actualinc

led by @Tom_A_Lynch

for the first time in the AI movement

open models are now good enough for a large and growing ially where latency, privacy, bandwidth, and token cost are critical

hermes agent from @nousresearch and ‘lab’ by @primeintellect have broken this idea wide open with killer products today, built on open rails

we are no longer forced to assume that every useful task must route through one of a few centralized frontier APIs

and so

tom and the squad at @actualinc have painstakingly created a new type of proprietary inference engine

for all types of computers

maybe you are an owner of one of the tens of millions Apple macs

or an owner of one of the millions of RTX-class machines

or maybe you run a small team or company

and you are starting to realize

what many others across the world, have started to realize

which is that sometimes centralized AI systems bring their own set of unique problems

especially as our workflows move into an always-on, agentic era which might sometimes come with:

- exorbitant costs
- no privacy
- lack of control

but a problem remains

running local AI is still too hard

firing up one model on one machine is difficult enough

but running models across different machines means navigating drivers, networking, model sizing, memory limits, API setups

stitching together hardware is messy across Mac, Windows, Linux, NVIDIA, Apple, AMD

heterogenous systems were never meant to work seamlessly together like this

today, you'll find that most of the local inference software available is built around single users, for a single machine use case

running inference capacity in bulk, across many users, and many kinds of computers

that is a different animal altogether

tools like llama.cpp made local models accessible, but they are not designed to coordinate tens of thousands of heterogeneous machines under a live supply network

vLLM-style batching can be economical, but vLLM is limited in the machines it can run on

the local AI movement requires a simple product experience, for a tough engineering problem

that is the problem @Tom_A_Lynch pitched me fifteen months ago

and the problem @actualinc has been engineered to solve

now with the help of his cracked and growing team @skooookum, @somewheresy, @jack15032, and the rest of the crew

starting with their first product

go to actual.computer

in under one minute, install the @actualinc client, authorize your machine, load a model, and @actualinc turns your personal device into a private inference endpoint

in less than twenty seconds, add a second, or third, or fourth machine of any type and it will start behaving like a small private cluster

from a technical POV

actual runs as a lightweight daemon on each machine and turns local hardware into a routable inference endpoint

each node reports its available memory, accelerator type, model state, and health back into the cluster, while the user interacts with a single API surface

instead of managing ports, drivers, or SSH sessions
a relay layer lets requests reach machines without forcing users to configure networking or expose their devices directly

and as a result

computer owners of all kinds of computers

can get local inference clusters

- when they want to control costs
- when they want to control privacy
- when they want to build on top of a neutral stack

eventually, over time

computer owners running these local clusters should be able to seamlessly contribute their unused capacity back to the @actualinc network

for a market to price

and for someone else to purchase

you can think of @actualinc as building the engine and control plane for datacenter style throughput in homes, in offices, in small to medium sized businesses, and in very big sized businesses

because there are hundreds of gigawatts already on the grid right now, that are waiting

on desks, in storage closets

to become actual computers

Actual Computer is pleased to announce its first source available-software, toks.

Actual's toks is the best tokenizer on earth and the first tokenizer designed for the quadrillion-token era.

actual.inc/company/blog/i…

We release toks under the source-available BUSL 1.1.

Free for any organization that processes fewer than one quadrillion tokens a year. For those above the mark, we offer an inexpensive license.

Get it now: github.com/actual-compute…

Toks is now released under Apache 2.0