https://www.tplr.ai/_next/static/chunks/368-2a930488df3b1f6b.js?dpl=dpl_E14WBDSVgBhEAh4maovWu6JCAqvj

1(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[368],{2773:(e,i,a)=>{"use strict";a.d(i,{SN:()=>n,_L:()=>t,XA:()=>r}),a(7345);let r=[{slug:"skipping-stages-with-fixed-projections",kind:"blog",internal:!0,reportNumber:"BLOG-04",label:"Templar Research Blog",date:"September 2026",datePublished:"2026-09-16",title:"Fault tolerance in low-bandwidth model parallelism: exploring pipeline stage-skipping with boundary compression.",shortTitle:"Fault tolerance in low-bandwidth model parallelism: exploring pipeline stage-skipping with boundary compression.",authors:["Templar Research"],deck:"We explore how the residual nature of the transformer architecture can be leveraged to mitigate hardware faults, and demonstrate that the compression-based implementation of low-bandwidth model parallelism within Crucible increases robustness to stage failures.",thumbnail:"/images/blog/fault-tolerance/pipeline-light.svg",thumbnailDark:"/images/blog/fault-tolerance/pipeline.svg",thumbnailAlt:"A failed inner stage is bypassed for multiple steps while the other stages continue training."},{slug:"introducing-crucible",kind:"blog",internal:!0,reportNumber:"BLOG-03",label:"Templar Research Blog",date:"August 2026",datePublished:"2026-08-24",readTime:"16 min read",title:"Introducing Crucible and An Economic Validation of Globally Distributed Pre-Training",shortTitle:"Introducing Crucible",authors:["Templar Research"],deck:"Our new scalable pre-training platform processed 50B tokens through an 8B model over globally distributed GPUs at a cost of $0.1202 per million tokens, 29% below a centralized AWS A100 cluster.",thumbnail:"/images/blog/crucible/crucible-system-topology.png",thumbnailDark:"/images/blog/crucible/crucible-system-topology-dark.png",thumbnailAlt:"Crucible orchestrator and three distributed four-stage pipeline replicas coordinating through object storage, with variable low-bandwidth links between adjacent stages.",arxivUrl:"https://arxiv.org/pdf/2601.02360",headerLinks:[{label:"Dashboard: TEST_RUN_003",href:"/dashboard"},{label:"Read the paper",href:"https://arxiv.org/pdf/2601.02360"}]},{slug:"pulseloco",kind:"blog",reportNumber:"BLOG-02",label:"Templar Research Blog",date:"May 2026",title:"Introducing PULSELoCo: More Than 100x Lower Trainer Communication for Distributed RL",shortTitle:"PULSELoCo",authors:["Covenant Labs","Erfan Miahi"],deck:"PULSELoCo applies the PULSE compute-visibility rule to DiLoCo-style pseudo-gradients: it matches DiLoCo learning behavior while sending over 17x less data than DiLoCo and about 138x less than DDP on Qwen2.5-7B.",thumbnail:"/images/blog/pulseloco/bandwidth-utilization.png",thumbnailAlt:"Comparison of trainer bandwidth requirements and bandwidth utilization for distributed reinforcement learning methods.",substackUrl:"https://templarresearch.substack.com/p/introducing-pulseloco-more-than-100x",arxivUrl:"https://arxiv.org/pdf/2602.03839"},{slug:"covenant-72b",reportNumber:"TPLR-005",label:"Templar Technical Report",date:"March 2026",title:"Covenant-72B: Pre-Training a 72B LLM with Trustless Peers Over-the-Internet",shortTitle:"Covenant-72B",previewKind:"covenant",authors:["Joel Lidin","Amir Sarfi","Erfan Miahi","Quentin Anthony","Shivam Chauhan","Evangelos Pappas","Benjamin Therien","Eugene Belilovsky","Samuel Dare"],deck:"A 72B-parameter language model trained with open, permissionless participation over commodity internet links, using Gauntlet incentives and SparseLoCo compression to keep unreliable peers useful.",summary:"Permissionless decentralized pre-training reaches the 72B scale while remaining competitive with centralized baselines.",arxivUrl:"https://arxiv.org/pdf/2603.08163",alphaxivUrl:"https://www.alphaxiv.org/abs/2603.08163",sections:[{id:"overview",label:"Overview"},{id:"method",label:"Method"},{id:"systems",label:"Systems"},{id:"results",label:"Results"},{id:"future-work",label:"Future Work"},{id:"references",label:"References"},{id:"citation",label:"Citation"}],bibtex:"@misc{lidin2026covenant72b,\n  title={Covenant-72B: Pre-Training a 72B LLM with Trustless Peers Over-the-Internet},\n  author={Lidin, Joel and Sarfi, Amir and Miahi, Erfan and Anthony, Quentin and Chauhan, Shivam and Pappas, Evangelos and Therien, Benjamin and Belilovsky, Eugene and Dare, Samuel},\n  year={2026},\n  eprint={2603.08163},\n  archivePrefix={arXiv},\n  primaryClass={cs.LG},\n  url={https://arxiv.org/abs/2603.08163}\n}"},{slug:"pulse",reportNumber:"TPLR-004",label:"Templar Technical Report",date:"February 2026",title:"Understanding and Exploiting Weight Update Sparsity for Communication-Efficient Distributed RL",shortTitle:"PULSE",previewKind:"pulse",authors:["Erfan Miahi","Eugene Belilovsky"],deck:"RL fine-tuning changes only about 1% of model weights per step. PULSE exploits that sparsity to synchronize policy weights losslessly with over 100x less bandwidth.",summary:"Lossless sparse weight patches turn RL weight synchronization from a 20 Gbit/s problem into a commodity-network one.",arxivUrl:"https://arxiv.org/pdf/2602.03839",alphaxivUrl:"https://www.alphaxiv.org/abs/2602.03839",sections:[{id:"overview",label:"Overview"},{id:"sparsity",label:"Why Updates Are Sparse"},{id:"interactive",label:"Interactive Figure"},{id:"method",label:"The PULSE Method"},{id:"results",label:"Results"},{id:"references",label:"References"},{id:"citation",label:"Citation"}],bibtex:"@misc{miahi2026pulse,\n  title={Understanding and Exploiting Weight Update Sparsity for Communication-Efficient Distributed RL},\n  author={Miahi, Erfan and Belilovsky, Eugene},\n  year={2026},\n  eprint={2602.03839},\n  archivePrefix={arXiv},\n  primaryClass={cs.LG},\n  url={https://arxiv.org/abs/2602.03839}\n}"},{slug:"heterogeneous-pretraining",reportNumber:"TPLR-003",label:"Templar Technical Report",date:"January 2026",title:"Heterogeneous Low-Bandwidth Pre-Training of LLMs",shortTitle:"Heterogeneous Pre-Training",previewKind:"heterogeneous",authors:["Yazan Obeidi","Amir Sarfi","Joel Lidin","Paul Janson","Eugene Belilovsky"],deck:"A framework for letting resource-limited participants form a SparseLoCo replica through compressed pipeline parallelism while better-connected replicas stay full precision.",summary:"Selective activation compression lets weak links contribute without forcing every replica into the same lossy regime.",arxivUrl:"https://arxiv.org/pdf/2601.02360",alphaxivUrl:"https://www.alphaxiv.org/abs/2601.02360",sections:[{id:"overview",label:"Overview"},{id:"topology",label:"Topology"},{id:"method",label:"Method"},{id:"results",label:"Results"},{id:"implications",label:"Implications"},{id:"references",label:"References"},{id:"citation",label:"Citation"}],bibtex:"@misc{obeidi2026heterogeneous,\n  title={Heterogeneous Low-Bandwidth Pre-Training of LLMs}

1,\n  author={Obeidi, Yazan and Sarfi, Amir and Lidin, Joel and Janson, Paul and Belilovsky, Eugene},\n  year={2026},\n  eprint={2601.02360},\n  archivePrefix={arXiv},\n  primaryClass={cs.LG},\n  url={https://arxiv.org/abs/2601.02360}\n}"},{slug:"grail-v0",kind:"blog",reportNumber:"BLOG-01",label:"Templar Research Blog",date:"December 2025",title:"grail-v0: How We Built a Fully Open, Incentivized, Decentralized Reinforcement Learning System",shortTitle:"grail-v0",authors:["Covenant Labs","Erfan Miahi"],deck:"The architecture, verification pipeline, and incentives behind one of the first open decentralized RL networks with real-time transparency — miners generate rollouts, validators verify them with the grail proof, and a trainer learns from a live network.",thumbnail:"/images/blog/archive/grail-v0/figure-01-grail-v0-architecture.png",thumbnailAlt:"grail-v0 architecture with miners, validators, and a live trainer.",substackUrl:"https://templarresearch.substack.com/p/grail-v0-how-we-built-a-fully-open"},{slug:"sparseloco",reportNumber:"TPLR-002",label:"Templar Technical Report",date:"August 2025",title:"Communication Efficient LLM Pre-Training with SparseLoCo",shortTitle:"SparseLoCo",previewKind:"sparseloco",authors:["Amir Sarfi","Benjamin Therien","Joel Lidin","Eugene Belilovsky"],deck:"SparseLoCo replaces dense pseudo-gradient exchange with chunked TOP-k sparsification, local error feedback, and 2-bit values while matching or improving DiLoCo-style pre-training.",summary:"A sparse, quantized local optimizer moves the loss-communication frontier for internet-scale pre-training.",arxivUrl:"https://arxiv.org/pdf/2508.15706",alphaxivUrl:"https://www.alphaxiv.org/abs/2508.15706",sections:[{id:"overview",label:"Overview"},{id:"intuition",label:"Intuition"},{id:"algorithm",label:"Algorithm"},{id:"results",label:"Results"},{id:"deployment",label:"Deployment"},{id:"references",label:"References"},{id:"citation",label:"Citation"}],bibtex:"@misc{sarfi2025sparseloco,\n  title={Communication Efficient LLM Pre-Training with SparseLoCo},\n  author={Sarfi, Amir and Therien, Benjamin and Lidin, Joel and Belilovsky, Eugene},\n  year={2025},\n  eprint={2508.15706},\n  archivePrefix={arXiv},\n  primaryClass={cs.LG},\n  url={https://arxiv.org/abs/2508.15706}\n}"},{slug:"gauntlet",reportNumber:"TPLR-001",label:"Templar Technical Report",date:"May 2025",title:"Incentivizing Permissionless Distributed Learning of LLMs",shortTitle:"Gauntlet",previewKind:"gauntlet",authors:["Joel Lidin","Amir Sarfi","Evangelos Pappas","Samuel Dare","Eugene Belilovsky","Jacob Steeves"],deck:"Gauntlet scores pseudo-gradient contributions from untrusted peers with fast reliability checks, loss-before/after validation, and persistent OpenSkill ratings.",summary:"A blockchain-coordinated incentive mechanism turns permissionless pseudo-gradients into useful training signal.",arxivUrl:"https://arxiv.org/pdf/2505.21684",alphaxivUrl:"https://www.alphaxiv.org/abs/2505.21684",sections:[{id:"overview",label:"Overview"},{id:"problem",label:"Problem"},{id:"mechanism",label:"Mechanism"},{id:"scoring",label:"Scoring"},{id:"results",label:"Results"},{id:"references",label:"References"},{id:"citation",label:"Citation"}],bibtex:"@misc{lidin2025gauntlet,\n  title={Incentivizing Permissionless Distributed Learning of LLMs},\n  author={Lidin, Joel and Sarfi, Amir and Pappas, Evangelos and Dare, Samuel and Belilovsky, Eugene and Steeves, Jacob},\n  year={2025},\n  eprint={2505.21684},\n  archivePrefix={arXiv},\n  primaryClass={cs.LG},\n  url={https://arxiv.org/abs/2505.21684}\n}"}].filter(e=>!1!==e.published),t=r.filter(e=>"blog"!==e.kind),n=r.filter(e=>"blog"===e.kind);r.filter(e=>"blog"===e.kind&&e.internal)},3780:e=>{e.exports={page:"publications_page__sPiHV",main:"publications_main__9HNqy",toolbar:"publications_toolbar__pMSfC",viewSwitch:"publications_viewSwitch___0Aqz",viewTab:"publications_viewTab__5GE7w",active:"publications_active__TSymr",masthead:"publications_masthead__cIhZk",mastheadMeta:"publications_mastheadMeta__pI_FJ",rowMeta:"publications_rowMeta__5EEzP",rowByline:"publications_rowByline__M_lAf",rowLinks:"publications_rowLinks__yF6iW",postList:"publications_postList__dyQh2",postRow:"publications_postRow__OTeVY",thumbnail:"publications_thumbnail__oS8yE",noThumb:"publications_noThumb__2c6RZ",rowCopy:"publications_rowCopy__lqFGC",vacancyRow:"publications_vacancyRow__M4JJo",vacantAuthor:"publications_vacantAuthor__7oy0X",caretBlink:"publications_caretBlink__AXdf5"}},4897:(e,i,a)=>{"use strict";a.d(i,{_:()=>r,z:()=>t});let r=e=>e.internal?"/publications/blog/".concat(e.slug):"blog"===e.kind?e.substackUrl:e.arxivUrl,t=(e,i)=>{let a=r(e);

1return a?a.startsWith("/")?"".concat(i).concat(a):a:i}},7257:(e,i,a)=>{"use strict";a.d(i,{A:()=>h});var r=a(3469),t=a(3499),n=a.n(t),l=a(1623),s=a(2773),o=a(3780),c=a.n(o);let d=e=>String(e).padStart(2,"0");function h(e){let{active:i}=e;return(0,r.jsxs)("nav",{className:c().viewSwitch,"aria-label":"Publication type",children:[(0,r.jsxs)(n(),{className:(0,l.A)(c().viewTab,"publications"===i&&c().active),href:"/publications","aria-current":"publications"===i?"page":void 0,children:[(0,r.jsx)("span",{children:"Papers"}),(0,r.jsx)("span",{children:d(s._L.length)})]}),(0,r.jsxs)(n(),{className:(0,l.A)(c().viewTab,"blog"===i&&c().active),href:"/publications/blog","aria-current":"blog"===i?"page":void 0,children:[(0,r.jsx)("span",{children:"Blog"}),(0,r.jsx)("span",{children:d(s.SN.length)})]})]})}},7782:(e,i,a)=>{"use strict";a.d(i,{A:()=>u});var r=a(3469),t=a(3499),n=a.n(t),l=a(1623),s=a(4349),o=a(2773),c=a(4897),d=a(3780),h=a.n(d);let p=e=>{let{article:i,href:a,className:t,children:l,...s}=e;return i.internal?(0,r.jsx)(n(),{href:a,className:t,...s,children:l}):(0,r.jsx)("a",{href:a,className:t,target:"_blank",rel:"noopener noreferrer",...s,children:l})};function u(e){let{articles:i,label:a,showVacancy:t=!1}=e;return(0,r.jsxs)("section",{className:h().postList,"aria-label":a,children:[i.map(e=>{let i=(0,c._)(e);return(0,r.jsxs)("article",{className:(0,l.A)(h().postRow,!e.thumbnail&&h().noThumb),children:[(0,r.jsxs)("div",{className:h().rowMeta,children:[(0,r.jsx)("span",{children:e.date}),(0,r.jsx)("span",{children:e.reportNumber})]}),(0,r.jsxs)("div",{className:h().rowCopy,children:[(0,r.jsx)("h2",{children:(0,r.jsx)(p,{article:e,href:i,children:e.title})}),(0,r.jsx)("div",{className:h().rowByline,children:(0,r.jsx)("span",{children:e.authors.join(" / ")})}),(0,r.jsx)("p",{children:e.deck})]}),e.thumbnail&&(0,r.jsx)(p,{article:e,href:i,className:h().thumbnail,"aria-label":"Read ".concat(e.title),children:(0,r.jsx)(s.M,{article:e})})]},e.slug)}),t&&(0,r.jsxs)("article",{className:(0,l.A)(h().postRow,h().noThumb,h().vacancyRow),children:[(0,r.jsxs)("div",{className:h().rowMeta,children:[(0,r.jsx)("span",{children:"TBD"}),(0,r.jsx)("span",{children:(()=>{let e=o.XA.reduce((e,i)=>{let a=/^TPLR-(\d+)$/.exec(i.reportNumber);return a?Math.max(e,Number(a[1])):e},0);return"TPLR-".concat(String(e+1).padStart(3,"0"))})()})]}),(0,r.jsxs)("div",{className:h().rowCopy,children:[(0,r.jsx)("h2",{children:(0,r.jsx)(n(),{href:"/careers",children:"This report doesn't exist yet. Write it with us."})}),(0,r.jsxs)("div",{className:h().rowByline,children:[(0,r.jsx)("span",{className:h().vacantAuthor,children:"Your Name Here"}),(0,r.jsx)("span",{children:"Templar Research"})]}),(0,r.jsx)("p",{children:"Every paper above started as an open problem. We hire researchers and engineers to work on the next one: optimizers, incentives, and systems for training frontier models over the open internet. We also collaborate with external researchers and labs, and offer research compute grants for promising work on decentralized intelligence."}),(0,r.jsxs)("div",{className:h().rowLinks,children:[(0,r.jsx)(n(),{href:"/careers",children:"Open Roles / Careers"}),(0,r.jsx)("a",{href:"mailto:[email protected]?subject=Research%20Collaboration%20%2F%20Compute%20Grant",children:"Collabs & Compute Grants"}),(0,r.jsx)("a",{href:"https://discord.gg/VUgfADep",target:"_blank",rel:"noopener noreferrer",children:"Say Hello / Discord"})]})]})]})]})}}}]);