The 118-billio -pa amete model matches models seve al times its size o age tic codi g a d is small e ough to u o a si gle desktop. It is available today u de a Ope MDW-1.1 lice se.
SAN FRANCISCO, July 21, 2026 (GLOBE NEWSWIRE) — Poolside today eleased
Lagu a S 2.1 a ives as the questio of who supplies the West’s ope -weight models has moved f om esea ch ci cles i to boa d ooms a d Washi gto . Ove the past yea , develope usage has shifted sha ply towa d ope -weight systems that teams ca dow load, i spect, a d u o thei ow i f ast uctu e, a d most of the leadi g optio s i that catego y ow come f om Chi ese labs. As seve al US labs have moved thei most capable systems behi d closed APIs a d gove me t-gated ollouts, the West has bee left without a competitive ope -weight optio . Lagu a S 2.1 is built to be that optio , a d to compete with the i cumbe ts o the te ms buye s ow ca e about: cost, co t ol, a d the ability to keep data i house.
The elease also ef ames a debate that has ce te ed o f o tie capability a d the capital equi ed to each it. As e te p ises move f om u co st ai ed toke spe di g towa d a focus o efficie cy a d etu , the ope ative questio s have become the p ice of u i g i tellige ce at scale, whethe that i tellige ce ca be self-hosted fo sove eig ty a d complia ce, a d whethe the busi ess behi d a model is sustai able. Lagu a S 2.1 is desig ed fo exactly those co st ai ts.
Poolside is eleasi g Lagu a S 2.1 alo gside a key set of age tic codi g evaluatio s u usi g pool, Poolside’s age t ha ess, so esea che s a d develope s ca ep oduce the esults o thei ow ha dwa e.
| Te ce t Hy3 | 295B-A21B | 71.7% | 75.8% | 57.9% | – | Thi ki g Machi es I kli g | 975B-A41B | 63.8% | – | 54.3% | – | Nemot o 3 Ult a | 550B-A55B | 56.4% | 67.7% | – | – | Mi iMax M3 | 428B-A23B | 66.0% | – | 59.0% | – | DeepSeek V4-Flash Max | 284B-A13B | 61.8%** | 73.3% | 52.6% | – | DeepSeek V4-P o Max | 1.6T-A49B | 64.0%** | 76.2% | 55.4% | 9.0% | Qwe 3.7-Max | – | 74.5%** | 78.3% | 60.6% | – | Claude So et 5 | – | 80.4% | – | 63.2% | 54.0% | GPT-5.6 Lu a Max | – | 82.5% | – | – | 67.0% |
Compa iso sco es use the highest publicly efe e ced esults fo each model. Full methodology, sampli g pa amete s, a d pe -be chma k detail will accompa y the elease.
** As epo ted by A tificial A alysis.
As a fu the demo st atio of lo g-ho izo easo i g, Poolside publishes a complete t ajecto y of Lagu a S 2.1 i depe de tly p ovi g E dős P oblem #397, a combi ato ics p oblem that u til ece tly o ly the la gest f o tie easo i g models had esolved.
Ru i g Lagu a S 2.1 locally cha ges the a ithmetic of deployi g codi g age ts. Small e ough to u o a si gle DGX Spa k, it lets teams move high-volume age tic wo k off mete ed APIs a d o to ha dwa e they co t ol. Fo the gove me t, defe se, a d othe highly egulated o ga izatio s that Poolside al eady se ves, self-hosti g keeps se sitive code a d data i side thei ow e vi o me t. Lagu a S 2.1 also occupies a size class that o Weste lab has eleased a ope -weight model i to i 11 mo ths, si ce gpt-oss-120b’ elease i August last yea .
Lagu a S 2.1 is the latest model out of Poolside’s Model Facto y, the compa y’s i te al platfo m fo t ai i g, evaluati g, a d ite ati g o fou datio models. The Facto y automates the wo k that t aditio ally makes model developme t slow, f om a chitectu e ablatio s to ei fo ceme t lea i g f om code executio , u i g expe ime ts that o ce took weeks i u de a hou . It is what lets Poolside ship o a oughly five-week cade ce. Fo efe e ce, Lagu a S 2.1 took less tha 4 weeks to t ai e d to e d o 4,000 H200 GPUs.
“The West eeds ope -weight models it ca t ust, u , a d build o ,” said Jaso Wa e , co-CEO of Poolside. “Lagu a S 2.1 is ou a swe . It is a model that e te p ises a d gove me ts ca put i to p oductio today, o thei ow ha dwa e, at a cost that makes age tic codi g p actical at scale.”
“Lagu a S 2.1 does the wo k of models seve al times its size because of how we build, ot despite it,” said Eiso Ka t, co-fou de a d co-CEO of Poolside. “The Model Facto y automates the wo k that t aditio ally makes model developme t slow a d e o p o e, allowi g its esea che s to u a o de of mag itude mo e expe ime ts.”
Lagu a S 2.1 is available today. Weights a e o Huggi g Face u de a Ope MDW-1.1 lice se. The model is also available th ough Poolside’s API a d o Ope Route , a d u s i pool, Poolside’s age t ha ess.
Poolside is a fou datio AI lab buildi g ope -weight models fo softwa e e gi ee i g o the path to huma -level a tificial i tellige ce. Fou ded i 2023 a d headqua te ed i Sa F a cisco, Poolside t ai s its models f om sc atch o its ow data a d i f ast uctu e, usi g ei fo ceme t lea i g f om code executio feedback so its models imp ove f om thei ow successes a d failu es, the way huma develope s do. Poolside deploys di ectly i to the developme t e vi o me ts of gove me t, defe se, a d othe highly egulated o ga izatio s. Lea mo e at https://poolside.ai.
