August 5, 2026

Small model, big main character energy

Beating GPT-5.6 Sol on retrieval with 100x cheaper open models

Tiny cheap AI just tied a big-name rival, and the comments are losing it

TLDR: Castform says it trained a small open AI model to search company data as well as GPT-5.6 Sol for about 100 times less money, which could make business AI much cheaper. Commenters were split between cheering the rise of focused budget models and asking the awkward question: what happens when your company documents are a mess?

A tiny open model just walked into the room, matched a much pricier AI on search-style fact finding, and immediately started a comment-section riot. The big flex from Castform is simple enough for normal humans: a small, cheap model was trained to dig through company documents and find answers as well as GPT-5.6 Sol, while allegedly costing 100 times less. In a moment when many teams are groaning about slow, expensive AI tools, that claim landed like a reality-show plot twist.

The crowd reaction? Equal parts hype, suspicion, and nerdy gossip. One camp was thrilled by the idea of custom-built little helper models, with people basically saying, “Finally, not every problem needs a giant diva model.” Another commenter tossed in a side quest by praising another budget-friendly model for its shockingly good one-shot performance, which gave the whole thread a “the cheap kids are cooking” vibe.

But the skeptics showed up fast. The biggest drama was over whether this kind of system can actually find the deeply buried stuff in a mountain of documents, especially when the answer depends on connecting two hidden clues. Others raised the messiest real-world problem of all: what if a company’s own files are outdated, misleading, or flat-out wrong? In other words, if you train on office chaos, do you just get a faster chaos machine?

And then came the funniest hot take: maybe smaller models win because big ones overthink while the little guys “just do it.” Brutal, hilarious, and honestly the comment-section quote of the day.

Key Points

  • Castform says a 4B open-source model post-trained for retrieval matched GPT-5.6 Sol in search-result accuracy while costing about 100x less.
  • The article describes a shift from embedding-based RAG pipelines toward agentic, multi-hop retrieval workflows that require repeated model calls.
  • It states that a typical multi-turn search request with GPT-5.6 Sol takes more than 10 seconds and costs roughly $0.03 end to end.
  • Castform uses Neon’s Lakebase Postgres and Lakebase Search across corpus storage, synthetic data generation, RL training, and production inference.
  • The article argues that enterprises already hold useful proprietary data for training, and Castform converts that corpus into tasks and manages the RL training loop.

Hottest takes

"specialized LLMs are being developed" — mrinterweb
"larger models overthink it while smaller ones just do it" — JCharante
"outdated/misleading or just plain wrong" — richwater
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.