Hed a b i gs the self-imp oveme t loop behi d f o tie esea ch to obotics, ma ufactu i g, biology a d defe se teams. I u de a week, its team made a p oductio image model oughly 6x faste o the same ha dwa e.
Sa F a cisco, CA, Nov. 10, 2026 (GLOBE NEWSWIRE) — Hed a, a a16z-backed esea ch compa y buildi g models a d i f ast uctu e fo multimodal a d physical AI, today lau ched Hed a Model Facto y, a auto omous post-t ai i g a d i fe e ce platfo m fo buildi g models fo the eal wo ld. The platfo m combi es flexible GPU sa dboxes, post-t ai i g u by Hed a esea che s wo ki g with auto omous AI age ts, co-optimized low-late cy deployme t, a d a data flywheel that keeps imp ovi g each model afte it ships.
Michael Li gelbach, fou de a d CEO of Hed a.
Models that imp ove themselves a e accele ati g f o tie esea ch. Hed a Model Facto y applies that loop to eal-wo ld p oblems, whe e data is multimodal a d o de s of mag itude la ge tha text, a d whe e ei fo ceme t lea i g e vi o me ts a d simulato s ofte eed GPU accele atio just to u at scale. Hed a g ew eve ue 15x ove the past 12 mo ths, a d the platfo m ope s the esea ch a d i f ast uctu e behi d that g owth to teams that eed custom models built o thei ow data.
“Recu sive self-imp oveme t is al eady cha gi g how f o tie models get built,” said Michael Li gelbach, fou de a d CEO of Hed a. “The ext step is the eal wo ld, whe e the data is video, se so s a d simulatio . Hed a Model Facto y gives obotics, ma ufactu i g, biology a d defe se teams the i f ast uctu e to post-t ai , deploy a d keep imp ovi g thei ow models.”
F om Compute to a Data FlywheelCompute. Hed a GPU sa dboxes give teams flexible capacity u i g o Hed a’s i fe e ce e gi e, so they ca scale t ai i g up without locki g i to lo g-te m GPU commitme ts.
Post-t ai i g. Teams b i g thei simulatio e gi es a d data, a d Hed a helps build the ei fo ceme t lea i g e vi o me ts a d datasets they eed. Hed a esea che s wo k alo gside each custome ’s team, usi g tools such as visio -la guage models that captio a d sco e data o give feedback du i g o li e ei fo ceme t lea i g. Fo compa ies without thei ow esea ch lab, Hed a esea che s also set up data captu e, evaluatio s a d the t ai i g-to-se vi g loop.
Deploy. Models u o low-late cy compute, o p emises o at Hed a’s ea est pa t e data ce te , f om distilli g wo ld actio models to u faste tha eal time without losi g physical cohe e ce to fitti g a visio -la guage-actio model o to a edge device. Teams ca also b i g a existi g model to se ve o Hed a’s i fe e ce e gi e.
Self-imp ove. O ce a model is deployed, Hed a esea che s a alyze its p oductio pe fo ma ce, et ai the model a d elease o ly ve sio s that pass the custome ‘s evals.
Resea ch Results Behi d the Facto yPhysically co siste t video. Hed a et ai ed H3 fo physical co siste cy. The et ai ed model sco ed 34.9 o Google DeepMi d’s Physics-IQ be chma k, compa ed with 30.3 fo H3 Max, with compa able visual quality. A auto omously distilled ve sio e de ed a five-seco d clip i u de th ee seco ds o AMD MI355X GPUs.
Image editi g. Th ough distillatio a d i fe e ce optimizatio , Hed a cut 2K image edits f om about 22 seco ds to 3.6 seco ds o the same fou AMD I sti ct MI355X GPUs, oughly 6x faste , i u de a week. The baseli e a the o igi al bf16 model at 30 sampli g steps; the faste esult uses a distilled 14-step model. The model u s o MI355X o Hed a today.
Robot ha dwa e cha ges. I a simulatio study, Hed a fitted a obot u i g a t ai ed policy with a ew g ippe a d esto ed task success f om ea ze o to ea 100% with o et ai i g, by matchi g the ew g ippe ’s each a d g asp poi t i the w ist came a view.
Recu sive self-imp oveme t desc ibes systems that get bette at imp ovi g themselves. Hed a applies the idea to how models a e made: esea che s wo k alo gside auto omous AI scie tists that u expe ime ts a d ite ate o t ai i g u s, a d eve y deployed model feeds data back i to the ext ve sio . Hed a’s esea ch pushes towa d a model facto y that gets faste with eve y model it p oduces.
Li gelbach has spe t his ca ee o ei fo ceme t lea i g fo visual a d physical systems, fi st at Sta fo d a d NVIDIA. Hed a’s esea ch team b i gs togethe tale t f om Sta fo d, Be keley a d UIUC, with expe ie ce at NVIDIA, Meta Supe i tellige ce Labs, Fi ewo ks AI a d Salesfo ce, a d has published esea ch i ge e ative video a d embodied AI at Neu IPS, CVPR a d ICLR.
AvailabilityHed a Model Facto y is available sta ti g today at the ew hed a.com. Teams ca wo k with Hed a esea che s to build a d deploy a custom model fo the p oblem thei ge e al-pu pose model ca ‘t solve. They ca also sta t with flexible GPU sa dbox capacity without lo g-te m commitme ts o se ve a existi g model o Hed a’s i fe e ce e gi e.
About Hed aHed a is a a16z-backed esea ch compa y buildi g models a d i f ast uctu e fo multimodal a d physical AI. Lea mo e at hed a.com.
CONTACT: Fo mo e i fo matio , please co tact Eve ett Butle at Hed a o eve ett@hed a.com
