you can buy a whole dell workstation with an RTX Pro 6K WS inside of it for less than most places will sell you just the pro 6k

dario begged me in dms to not do agent swarms at home but whatever, this is nemotron 120B A12B running 64 concurrents at a bit over 100 tok/sec each for an aggregate of almost 7K tok/sec on a single GB300 (Asus DGX Station) believe it or not, the agents did NOT magically just start hacking the planet tho. weird.

wym? maybe I am misunderstanding you but they did include Opus 5.5 there.

Trying to train a fall recovery policy for the @asimovinc humanoid Getting up is hard!

Nemotron 3.5 Super VL (120B A12B) is now the leading US open source model. Congrats to the NVIDIA AI team! I am still doing testing of this model, but I think the ranking would be even stronger if we were just measuring agentic coding performance. This one may very well be a solid choice for a speedy local coding agent model I'm running many things in parallel, will get proper average and max numbers in time, but just randomly prodding this model for some quick speed checks and over 550 tok/sec. All vanilla settings. Once optimized 600+, maybe even 650+, will be trivial, and so far it does appear to be a token efficient problem solver. Really exciting to see what looks like a solid release so far from nvidia!

It is another good morning!

reminds me of when i kept my server rack in the kitchen

It's a good morning btw

Robot used here is @sourccey from Vulcan Robotics. If youre interested in an actually useful robot to R&D on like this, definitely hit them up. For a deeper dive into how all this is working at the moment: youtube.com/watch?v=dGkJe3aI…

No training, no sim2real gaps, just a prompt and a dream. The problem in robotics today is every task is a complex combination of locomotion, manipulation, vision, planning, and a bunch of other things that fall apart once you diverge from training data or characteristics from sim. This is why you see so many demos with perfect lighting conditions and always the same, or precisely placed objects, specifically chosen for the task and demo. You never see these things deployed in the real world despite seeing demos that look convincing, because everything falls flat in the real world with real unseen conditions. Every single deployment is a unique edgecase. Here, it's just a multimodal LLM (DSV4.1F) doing a task it was never deliberately trained for, using cameras it's never seen, servos/motors it didn't learn on, in an environment it's never dealt with and a form factor never tested. It is currently slow, this video is 10x speed, but I am very confident it can be sped up as is, but also this is basically the slowest this level of intelligence will ever be again. The more I work with general purpose multimodal LLMs, the more I am realizing that ACT, VLAs, and all this other stuff is probably just a sideshow and a distraction, and it's just going to largely get eaten up by general purpose models that just happen to also do robotics. I don't think it'll always be LLMs, but the fact that LLMs can do some of the most challenging robotics tasks tells me it's kind of over for a large class of robotics training. This does not include RL, however, for low level things, like gait, reaching, and other capabilities or "skills" for hardware. RL is still a king here and you really do want these models to be super fast and super small. RL is very good here and safe I'd say. Decision models like Clef, Jev...etc remain mostly useless imho for robotics. I look forward to someone proving out a useful recipe that works these things into robotics, but I don't currently see it.

That's ~ 63K tok/sec prefill and obviously 450 tok/sec decode is also lovely