--- title: The Prime Intellect Stack — Will Brown, Prime Intellect description: Will Brown from Prime Intellect presents their open-source AI post-training stack, covering the Verifiers V1 library and PrimeRL training framework. The talk... canonical: https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: The Prime Intellect Stack — Will Brown, Prime Intellect | daily.dev og:description: Will Brown from Prime Intellect presents their open-source AI post-training stack, covering the Verifiers V1 library and PrimeRL training framework. The talk... og:url: https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf og:image: https://api.daily.dev/og/posts/PJD7GuxyF.png og:image:alt: The Prime Intellect Stack — Will Brown, Prime Intellect og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # The Prime Intellect Stack — Will Brown, Prime Intellect **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 46 min read · 2 upvotes · 0 comments ## Summary Will Brown from Prime Intellect presents their open-source AI post-training stack, covering the Verifiers V1 library and PrimeRL training framework. The talk details a redesigned environment architecture decomposed into task sets, harnesses, and runtimes that support evaluation, SFT, RL, and on-policy distillation workflows. Key technical highlights include an interception server pattern that lets any agent harness work transparently with RL training, a trace graph data structure for managing message/token duality in agentic rollouts, a new 'renderers' library for robust chat template handling, and async RL design that decouples rollout speed from training throughput. Performance benchmarks show a full GLM-5 RL step on 28 nodes in under 5 minutes for 131K context coding tasks, enabling a 1,000-step frontier-model run for ~$50K. The platform supports multi-tenant LoRA training today with full fine-tuning coming soon. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is new in Verifiers V1 compared to the previous version of the verifiers library? Verifiers V1 is a full overhaul that decomposes environments into three composable pieces: a task set (data and agent-agnostic rules), a harness (the agent loop, tools, system prompts), and a runtime (where the harness executes, such as local subprocess, Docker, or Prime sandboxes). The old rubric pattern was removed in favor of a decorator-based, PyDantic-typed configuration system with TOML config support. The old multi-turn/tool environment pattern still works, but V1 is the new recommended approach. _Teams evaluating agent training frameworks can track library overhauls like this one on daily.dev._ ### How fast and how much does a large-scale RL training run on GLM-5 cost with Primer RL? A GLM-5 RL training step on 28 nodes completes in under 5 minutes for long-horizon coding tasks with 131K context, enabling a 1,000-step run in roughly 3 days at about 50,000 dollars in rental compute costs. This is presented as substantially cheaper than comparable frontier-scale training options from providers like OpenAI or dedicated GPU cluster vendors. _Engineers budgeting large RL runs can follow real-world cost benchmarks like this via daily.dev._ ### Why does async reinforcement learning work better than synchronous RL for training coding agents? Async RL avoids bottlenecking training on the slowest rollout, since coding agent tasks can range from 30 seconds to 3 hours long. The inference server always serves the latest model version while rollouts complete independently and feed into whatever batch is available, tolerating going meaningfully off-policy (an average of around 16 steps off in practice) without destabilizing training, using techniques like the DPPO loss. _Developers weighing sync versus async RL setups can dig into practical tradeoffs like these on daily.dev._ ## Similar posts on daily.dev - [How to Build an RL Environment](https://daily.dev/posts/how-to-build-an-rl-environment-u09fzsb0r) · Daily Dose of Data Science \| Avi Chawla \| Substack · 2 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"The Prime Intellect Stack — Will Brown, Prime Intellect","url":"https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf"},"datePublished":"2026-07-13T15:40:21.168Z","dateModified":"2026-09-14T07:12:43.734Z","description":"Will Brown from Prime Intellect presents their open-source AI post-training stack, covering the Verifiers V1 library and PrimeRL training framework. The talk...","image":"https://i.ytimg.com/vi/V-EDrhIhHzQ/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/V-EDrhIhHzQ/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,reinforcement-learning","timeRequired":"PT46M","video":{"@type":"VideoObject","name":"The Prime Intellect Stack — Will Brown, Prime Intellect","description":"Will Brown from Prime Intellect presents their open-source AI post-training stack, covering the Verifiers V1 library and PrimeRL training framework. The talk...","thumbnailUrl":"https://i.ytimg.com/vi/V-EDrhIhHzQ/sddefault.jpg","uploadDate":"2026-07-13T15:40:21.168Z","duration":"PT46M","url":"https://api.daily.dev/r/PJD7GuxyF","embedUrl":"https://www.youtube.com/embed/V-EDrhIhHzQ"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"The Prime Intellect Stack — Will Brown, Prime Intellect"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/the-prime-intellect-stack-will-brown-prime-intellect-pjd7guxyf#faq","mainEntity":[{"@type":"Question","name":"What is new in Verifiers V1 compared to the previous version of the verifiers library?","acceptedAnswer":{"@type":"Answer","text":"Verifiers V1 is a full overhaul that decomposes environments into three composable pieces: a task set (data and agent-agnostic rules), a harness (the agent loop, tools, system prompts), and a runtime (where the harness executes, such as local subprocess, Docker, or Prime sandboxes). The old rubric pattern was removed in favor of a decorator-based, PyDantic-typed configuration system with TOML config support. The old multi-turn/tool environment pattern still works, but V1 is the new recommended approach. Teams evaluating agent training frameworks can track library overhauls like this one on daily.dev."}},{"@type":"Question","name":"How fast and how much does a large-scale RL training run on GLM-5 cost with Primer RL?","acceptedAnswer":{"@type":"Answer","text":"A GLM-5 RL training step on 28 nodes completes in under 5 minutes for long-horizon coding tasks with 131K context, enabling a 1,000-step run in roughly 3 days at about 50,000 dollars in rental compute costs. This is presented as substantially cheaper than comparable frontier-scale training options from providers like OpenAI or dedicated GPU cluster vendors. Engineers budgeting large RL runs can follow real-world cost benchmarks like this via daily.dev."}},{"@type":"Question","name":"Why does async reinforcement learning work better than synchronous RL for training coding agents?","acceptedAnswer":{"@type":"Answer","text":"Async RL avoids bottlenecking training on the slowest rollout, since coding agent tasks can range from 30 seconds to 3 hours long. The inference server always serves the latest model version while rollouts complete independently and feed into whatever batch is available, tolerating going meaningfully off-policy (an average of around 16 steps off in practice) without destabilizing training, using techniques like the DPPO loss. Developers weighing sync versus async RL setups can dig into practical tradeoffs like these on daily.dev."}}]} ```