跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.26358 · v1

Beam Model Technical Report

Reflection TeamAdam AbdulhamidOsama AbuelsorourYasmin AlameddineCarlo AlfanoDuarte AlvesBrandon AmosSelvaraj AnandarajArseni AnisimovichGeorge AntoniadisIoannis AntonoglouSheila BabayanNeeraj BahlMärt BaklerVidhisha BalachandranCameron BanksThibaut BarroyerNavroop BathIrwan BelloJeremy BermanFernanda BernaldoMrityunjay BhadauriaSanyam BhutaniTolga BolukbasiAnudhyan BoralJoseph BrennanEthan BrooksSam BuchananAnton BulyenovJohn CaiCatherine CangEason ChanElijah ChanakiraNathan ChandrasekaranHao ChenHong-You ChenJeffrey ChenSaurav ChhatrapatiTing-Rui ChiangBrian Tiger ChowAakanksha ChowdheryTimothy ChungRob ClaireJohn D. Co-ReyesAnthony ConteEduardo Coronado-MontoyaJulien CrozonDrew CsillagJoe CummingsJeff DaSamira DarukiTim DawsonCharlie DeckArash DelijaniOusmane DiaRandy DodgenRyan DouglasOmar El-KishkyBen EllisGamaleldin ElsayedSimon FanPatrick FernandesGabriel FernsNick FramAnkush GargFlorent GbelidjiIaroslav GeraskinBehrooz GhorbaniIvan GlushkovNeil GollapudiScott GoodfriendYoni GozlanChristopher GrimmSuriya GunasekarShruti GuptaJoel GustafsonEmre GuvenJon HackmannJeremy HadidjojoJess HamrickEthan HellmanJoscha HocheSpurthi Amba HombaiahUzair HondaZijian HuDzmitry HubaNancy HungVladimir IvanovShomik JainMojan JavaheripiKevin Martin JoseAymen KallalaAmog KamsettyMin Hyung KangSteven KapturowskiRamtin KeramatiSamuel KesslerSabina KimSeonghyeon KimIvan KomarovAgni KumarThanard KurutachJessica LandonJames LarochelleMichael LaskinJonathan LeeJonathan LeungChenyu LiHaoran LiRongzhao LiMichelle LightBoyan LinHai LinXiayue Charles LinMichael LindenOllie LiuWei LiuZichang LiuBrian LuiAmr MahdiRaunak ManjaniKobe MartinLucas MaystreFrancis McCannBrendan McLaughlinIsabel MehtaAlishah MominGerard Moreno-Torres BertranAdriana MurgaSwarnashree Mysore SathyendraAshley NagakiVishal NandavanamSwaroop NathJake NationsAngus NgThien NguyenTyler NicholsAnaelia OvalleJames PellyMichelle PieszkoSashank PisupatiOleksandr PolozovPiyush PrasadDrew ProudSalman QuaziAsfandyar QureshiMarko RadmilacMohit RaghavendraVikash RaiRalph RectoAnurag ReddyAksel ReediAbby ReiderDavid RenardyMatthias ResoPierre RouxDavid RubinsteinPaul RuhIldar SafarovHimanshu SahniAdrian SamatanSimon SandrewAdriel SaportaRohan SavargaonkarVlad SavinovNick SchankJonathon SchwartzTim SeahAhmad SharifMiriam ShlimovichHamid ShojanazeriYusuf SindiTeak SittipuntKurt SmithNicolas SonneratJoe SpisakRaunak SrivastavaRichard SteigerwaldAdam SteinStu StewartArnaud StieglerJingtong SuAlicia SunZoey SunJakub SygnowskiDouglas TeixeiraChetan TekurZachary TellmanHelen ThatcherRomal ThoppilanAnil TolwaniKonstantin Topaloglou-MundyDmitry UlyumdzhievMeet VadakkancheryChristopher VanderschuereBalaji VaradarajanRohan VarmaBalaji VeeramaniAndreas VerdelisRisto VuorioTom WalsheAllen WangEwen WangTaiyi WangZach WentzPeter WieseTravis WilsonJason WintersSam WisemanArvin WuJin WuKai WuKenneth XuMira YadavXianjun YangYao-Yuan YangXiaolan YouRenault YoungChristina YuBiao ZhangGavin ZhangMingrui ZhangShenao ZhangJiesi ZhaoMengjun ZhaoXingchen ZhaoMike ZhouTong ZhouZhenpeng ZhouTina Zhu
首次提交 2026-11-20 00:28
版本更新 2026-11-20 00:28
雷达观测 2026-11-20 12:52
arXiv ↗PDF ↗

摘要

Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads. Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter (DFlash 2), Beam maximizes serving throughput with efficient, flexible deployment. Averaged across conversation turns and harm categories, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless requests, achieving the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-20 00:28 · 最新观测版本

Beam Model Technical Report

Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads. Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter (DFlash 2), Beam maximizes serving throughput with efficient, flexible deployment. Averaged across conversation turns and harm categories, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless requests, achieving the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier.

查看此版本解读 →查看此版本原文 ↗