Ornith Is the Open-Source Coding Model Built for Agents, Not Humans

DeepReinforce, an AI analysis lab beforehand recognized for CUDA-L1 and the IterX code-agent optimization loop, launched Ornith-1.0 late final week—a household of open-source coding fashions obtainable on Hugging Face in 4 sizes primarily based on the variety of parameters: 9 billion, 31 billion, 35 billion combination of specialists, and a 397 billion mixture-of-experts flagship, all below MIT license with no regional restrictions.

Parameters are mainly the variety of dials and configurations a mannequin can deal with on its coaching. The extra parameters, the extra succesful a mannequin is. A 9-billion-parameter mannequin is taken into account small, ok to run on a very good smartphone, however not able to doing any heavy reasoning job reliably. A 397 billion mannequin is way more succesful, however requires some heavy computing, the sort that isn’t obtainable on shopper {hardware}.

The lab describes it as “a self-improving household of open-source fashions specifically for agentic coding duties.” That phrase—agentic—is doing quite a lot of work.

Most AI that folks work together with is conversational: you kind, it responds, the alternate ends. Agentic AI is completely different—it will get a job and takes actions to finish it with out a human guiding every step. In a coding context, which means an AI that reads information, runs checks, identifies what failed, fixes the code, and loops once more till it is executed.

So Agentic AI means nobody must be on the keyboard for more often than not. That is the entire level. That is additionally the path the place essentially the most commercially related progress is going on in 2026—the fashions that may run unsupervised by way of 20-step dev workflows are price greater than those that write a clear operate on request.



Nevertheless, most giant language fashions are nonetheless designed with human suggestions in thoughts.

How Ornith’s mind works

Most AI coding brokers are paired with a human-designed harness—a set algorithm for a way the agent buildings its work: when to name a software, how one can deal with an error, how one can decompose a multi-step downside. Ornith as an alternative “treats the scaffold as a learnable object that co-evolves with the coverage.”

Translation: as an alternative of inheriting another person’s playbook, it develops its personal.

Throughout reinforcement studying, every coaching step occurs in two phases. The mannequin first reads the duty and proposes a refined technique for approaching it. Then it makes use of that technique to generate an answer.

The reward from the result flows again to each phases—so the mannequin is optimized for writing higher methods, not simply higher code. Try this 1000’s and hundreds of thousands of instances, and task-specific approaches emerge with out a human engineering them.

DeepReinforce additionally takes reward hacking significantly. If the mannequin can write its personal coaching scaffold, it may well theoretically write a scaffold that video games the verifier—touching a file to make it seem like it accomplished a job with out truly doing the work. Three layers of protection block this: the atmosphere and check suite are immutable and out of doors the mannequin’s attain, a deterministic monitor flags any try to entry restricted paths or alter verification scripts, and a frozen choose mannequin sits on prime of the automated verifier as a veto.

The numbers

The flagship 397 billion parameter mannequin posts 82.4 on SWE-bench Verified—a check the place an AI is given an actual bug from an open-source GitHub repository and should repair it with out seeing the check suite, scored as the share of points it efficiently resolves.

That beats Claude Opus 4.7’s 80.8 and DeepSeek-V4-Professional’s 80.6 on the identical check. On Terminal Bench 2.1—89 duties run inside containerized terminal environments starting from debugging async code to resolving safety vulnerabilities, scored by completion charge—it posts 77.5 towards Claude Opus 4.7’s 70.3. 

On condition that SWE-bench contamination issues have been raised publicly—OpenAI argued earlier this yr that fashions had been inflating scores by memorizing benchmark options seen throughout coaching—Ornith additionally studies numbers on SWE-bench Professional, a more durable model utilizing extra various, less-leaked codebases scored the identical manner. The 397 billion mannequin lands at 62.2 there. Meaningfully decrease, however nonetheless aggressive with the sphere, and nonetheless higher than Deepseek V4 Professional.

The 9 billion parameter mannequin could be the extra fascinating information level. It posts 69.4 on SWE-bench Verified—greater than Gemma 4-31B’s 52 and aggressive with Qwen 3.5-35B’s 70, regardless of being 3-4 instances smaller.

Who it is for, and who it is not

Ornith-1.0 is explicitly not a general-purpose AI. The mannequin’s personal documentation says it might underperform on duties outdoors agentic coding. In order for you AI to summarize a doc, allow you to write your doctoral thesis, or draft an electronic mail, Ornith-1.0 is the improper decide.

It is optimized for a slim downside set: developer pipelines the place an AI agent takes a job description, operates inside a code repository or terminal session, and completes multi-step work with out intervention. This can be a software that was constructed for people who find themselves already working agent infrastructure—not for individuals attempting to resolve if AI is price utilizing.

The “beats Claude” headline is actual however requires context. As Decrypt reported, each lab is now chasing efficiency on agentic coding evals, as a result of that is the place the helpful efficiency variations dwell.

Ornith-1.0-397B does surpass Claude Opus 4.7 on each completely different coding benchmarks, however Anthropic’s present flagship, Claude Opus 4.8, scores greater. The comparability that holds is throughout the open-source class, at comparable parameter counts, on coding-specific agent duties.

For builders constructing self-hosted coding pipelines, agentic infrastructure, or comparable coding-focused work, the small and medium fashions working on edge {hardware} could also be genuinely helpful, however the common Joe could also be higher wanting some place else.