SERA - Fast, accessible coding agents that adapt to any repo
SERA is a family of open coding models (8B, 14B, 32B) trained with a new efficient method. SERA learns from "soft-verified" data, drastically reducing training costs. Easily adaptable to private codebases. Open weights, data & recipes.
Replies
Flowtica Scribe
Hi everyone!
SERA (Soft-verified Efficient Repository Agents) is the latest from Ai2's Open Coding Agents project. They just updated it with a new 14B model and refreshed datasets.
Two technical points make this approach interesting:
First, SERA proves that models can learn effectively from "partially correct" patches—much like humans learning through debugging. This insight pushes synthetic data costs down significantly, with entry-level reproduction costing just ~$400.
Second, for teams with private codebases or specific internal frameworks, this is a solid option. You can specialize these models to your own stack efficiently.
Since they released everything (weights, data, recipe), it is a great resource if you want to build custom agents!
This post, written by @tim_dettmers, covers the story behind building SERA.
Love this direction — learning from “imperfect” patches feels much closer to how real dev work happens, and the low cost to specialize for private repos is a big win for teams.
nice tool ! I need ot check that !
Spent half a day undoing an agent change in the wrong monorepo package. SERA being open source and built to adapt to any repo is a strong start. Does the CLI show files touched, commands run, and tests passed before I apply a patch? That's the difference.
Congrats on the release — impressive work. I’m curious from a standards/consistency perspective: when SERA adapts to a new repo, how do you ensure it follows stable patterns instead of drifting between different coding styles? Is there any way to define explicit rules or constraints the model must follow during generation?