--- title: How Castform + Neon Beats Frontier Models on Price and... description: Castform and Neon have partnered to enable RL post-training of small open-source models on proprietary enterprise data stored in Postgres. A 4B parameter model... canonical: https://daily.dev/posts/how-castform-neon-beats-frontier-models-on-price-and-efficiency-vrqxucyig twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How Castform + Neon Beats Frontier Models on Price and Efficiency | daily.dev og:description: Castform and Neon have partnered to enable RL post-training of small open-source models on proprietary enterprise data stored in Postgres. A 4B parameter model... og:url: https://daily.dev/posts/how-castform-neon-beats-frontier-models-on-price-and-efficiency-vrqxucyig og:image: https://api.daily.dev/og/posts/vRqXucyIg.png og:image:alt: How Castform + Neon Beats Frontier Models on Price and Efficiency 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. # How Castform + Neon Beats Frontier Models on Price and Efficiency **[Neon](https://daily.dev/sources/neontech)** · 6 min read · 0 upvotes · 0 comments ## Summary Castform and Neon have partnered to enable RL post-training of small open-source models on proprietary enterprise data stored in Postgres. A 4B parameter model post-trained with Castform matched GPT-5.6 Sol on retrieval accuracy at 100x lower cost. The pipeline uses Neon's Lakebase Search extensions for corpus storage, synthetic data generation, and inference. Castform automates the RL training loop — generating question-answer pairs from existing documents, defining reward functions, and managing rollouts — without requiring ML or GPU expertise. Neon's autoscaling and branching handle the bursty, parallel workloads of thousands of training rollouts while keeping costs low. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Community take How the wider developer community reacted, aggregated from 1 discussion and 73 comments across hackernews (as of 2026-08-06). **TL;DR:** Commenters are broadly enthusiastic about the idea of specialized, fine-tuned small models beating frontier models on narrow tasks at far lower cost, though several push back on benchmark transparency and ask for comparisons against cheaper alternatives like DeepSeek Flash or Luna. **Sentiment:** 55% positive · 35% mixed · 10% skeptical **The case for** - Specialized small models can outperform larger general-purpose ones on narrow retrieval tasks, partly because bigger models tend to 'overthink' or go off on tangents. - Once a fine-tuning pipeline is set up, rerunning it on top of new base models is relatively trivial, reducing ongoing maintenance burden. - The ROI case for even small accuracy gains is compelling at scale (customer support, fraud detection, ads). - Using proprietary enterprise data for post-training is a genuine moat that general frontier models can't easily replicate. **The pushback** - Benchmark transparency is lacking — no comparison against common retrieval benchmarks like BrowseComp, and the evaluation corpus (GitLab handbook) may be too clean to reflect real enterprise data quality. - Cheaper alternatives like DeepSeek Flash or Luna are not benchmarked against the custom model, making the '100x cheaper than GPT-5.6 Sol' claim hard to contextualize. - Outdated, contradictory, or low-quality corpus data poses a real challenge when deriving reward functions from the corpus itself. - Skeptics note that strong general models already handle many specialized tasks well, and the 'specialized model beats general model' story has historically underdelivered. **By community** - hackernews (positive): Broadly receptive to the specialized-model thesis, with active founder engagement addressing benchmark and methodology questions, though a vocal minority demands more rigorous, independent evaluation and cheaper-model comparisons. **Hottest debate:** Whether the benchmark is meaningful given the lack of comparison against much cheaper models (DeepSeek Flash, Luna) and the absence of standard retrieval benchmarks. **Open questions** - How does the fine-tuned 4B model perform against DeepSeek Flash or Luna, which are already far cheaper than GPT-5.6 Sol? - How well does the retrieval hold up on larger, messier, real-world enterprise corpora with contradictory or outdated documents? - How effective is the approach at finding 'paired needles' — cases where one retrieved fact is needed to unlock finding another? - What is the latency profile of the custom model compared to frontier models? **Highlights** > This feels like the database equivalent of "use the right data structure." We've spent two years assuming the biggest general-purpose model should do everything. It makes more sense for retrieval, reranking, reasoning, and generation to each have their own optimized model if the routing cost is negligible. > — [BedVibe\_Studios on hackernews · 1 comments](https://news.ycombinator.com/item?id=49188548) > > Claude code does some of this by handing off the "explore" agent work to haiku. That is not handing off to a specialized model, its just handing off to a lighter and interior model (compared to the parent model). That by itself can create issues like the lighter model not capturing all the data that the parent needs. The idea is that we get specialized models that are better then general purpose models. But its rare for a specialized model to beat a strong general model. There is a reason why we hear less about this idea of smaller expert models, because large strong models to the tasks just as good. And if the tasks is repetitive to the point that specialization is useful, you can get into a situation that your better off having a program written for that reputative nature, then delegating to other models. And then have the main strong model, deal with the (semi)cleaned up data. > — [benjiro29 on hackernews · 3 comments](https://news.ycombinator.com/item?id=49188740) > On what do you guys test the model. Its very dubious that there is no common retrieval benchmark such as browsecomp plus or similar tested. And what metric do you report? > — [breadislove on hackernews · 2 comments](https://news.ycombinator.com/item?id=49188008) > Nice, but there's no mention of how Luna or DSFlash perform on the same task? (Being 25x and 50x cheaper respectively.) Nor of how much faster their custom model performs? > — [andai on hackernews · 1 comments](https://news.ycombinator.com/item?id=49190885) > There is a more serious question in here that's not being answered. How effective is the retrieval in finding buried needles in larger and larger haystacks. And there's a correlary question, how effective could you be in finding paired needles in that haystack where you need to hold a needle to unlock finding another needle. > — [aliljet on hackernews · 3 comments](https://news.ycombinator.com/item?id=49187553) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49186762) · 229 points · 73 comments ## Similar posts on daily.dev - [The Neon backend is GA: a complete set of primitives so agents can build](https://daily.dev/posts/the-neon-backend-is-ga-a-complete-set-of-primitives-so-agents-can-build-ueg1c5mpg) · Neon · 7 upvotes · 0 comments - [How Shopify built a continual learning loop with PyTorch and vLLM – PyTorch](https://daily.dev/posts/how-shopify-built-a-continual-learning-loop-with-pytorch-and-vllm-pytorch-viqzi06yu) · PyTorch · 0 upvotes · 0 comments --- Tags: [#rag](https://daily.dev/tags/rag), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning), [#agentic-ai](https://daily.dev/tags/agentic-ai) [View this post on daily.dev](https://daily.dev/posts/how-castform-neon-beats-frontier-models-on-price-and-efficiency-vrqxucyig) ```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":"How Castform + Neon Beats Frontier Models on Price and Efficiency","url":"https://daily.dev/posts/how-castform-neon-beats-frontier-models-on-price-and-efficiency-vrqxucyig","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-castform-neon-beats-frontier-models-on-price-and-efficiency-vrqxucyig"},"datePublished":"2026-08-05T17:24:28.780Z","dateModified":"2026-08-06T02:31:33.432Z","description":"Castform and Neon have partnered to enable RL post-training of small open-source models on proprietary enterprise data stored in Postgres. A 4B parameter model...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4ce1fc8e427c07a455d48508fae7815c?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4ce1fc8e427c07a455d48508fae7815c?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Neon","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":"Neon","logo":"https://media.daily.dev/image/upload/s--65EDghou--/f_auto/v1724339118/logos/neontech","url":"https://daily.dev/sources/neontech"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-castform-neon-beats-frontier-models-on-price-and-efficiency-vrqxucyig","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"rag,reinforcement-learning,agentic-ai","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Neon","item":"https://daily.dev/sources/neontech"},{"@type":"ListItem","position":3,"name":"How Castform + Neon Beats Frontier Models on Price and Efficiency"}]} ```