Ornith-1.0: Open-Weight Coding Models From 9B to 397B Under an MIT License

A model family called Ornith-1.0 has appeared as open source under the MIT license, and the numbers published for it are aimed squarely at agentic coding. The reported scores are 82.4 on SWE-Bench Verified, 77.5 on Terminal-Bench 2.1, 62.2 on SWE-Bench Pro, 48.2 on NL2Repo and 77.1 on ClawEval, with a 41.2 on the SWE Atlas QnA set as well. The pitch is that open models have stopped trailing in coding and now lead, across a whole range of sizes, deployable locally, with a GGUF build included. These are the release's own claims, so treat the leaderboard talk with some care while looking at what the family actually offers.

Four sizes

Ornith isn't one model but four: a 9B dense, a 31B dense, a 35B mixture-of-experts and a 397B mixture-of-experts. They're post-trained on top of Gemma 4 and Qwen 3.5. The spread is unusual, since open releases tend to pick a small efficient size for volume or a huge one for raw power. Here every tier has a plain audience. The 9B runs on consumer GPUs for individual developers at no cost. The 31B suits a small team with a single server. The 35B MoE targets mid-sized projects where efficiency counts, and the 397B MoE is meant for enterprises that want private deployment with data kept in-house. It reads less like publishing a model and more like claiming a whole track.

What the scores measure

SWE-Bench Verified hands the model a real GitHub issue and asks it to find the relevant code, change it and make every unit test pass. It's real programming work, not puzzle drilling. A skilled human developer is often put at roughly 70 to 75 percent on such tasks, which is a loose industry consensus, and the 82.4 reported for Ornith would be the highest publicly checkable figure on the benchmark. Note that it's a score, and the release doesn't present it as a percentage.

SWE-Bench Pro involves edits across several files, refactors that cross modules and tangled dependencies, so 62.2 suggests it copes with medium-sized engineering tasks and not just trivial tweaks. Terminal-Bench 2.1 checks work inside a real terminal: moving around directories, grepping, editing configuration, debugging services and deploying apps, and 77.5 implies actual terminal use as opposed to a simulation of it. NL2Repo asks for a whole repository generated from a plain-language description, such as a task manager app. 48.2 looks modest until you consider the difficulty, and it's described as the best result among open models.

A caution for readers: benchmark results depend on the harness used to run them, and the same model can land far apart under different setups, so rerunning the weights on your own repositories is the only test that tells you how a given size behaves for your work. Still, taken together, the figures describe something closer to an engineer who works at a terminal, reads a codebase and finishes development tasks on its own than a chatbot that talks about code. And it comes under MIT, free for commercial use.

Letting the model improve its own tooling

The scores aren't the most interesting part. How they were reached is. AI coding tools have two layers: the scaffold, which handles planning, tool calls and context management, and the solution, which writes code, fixes bugs and runs tests. Teams have normally hand-built the scaffold with fixed rules and trained only the solution side. Ornith's approach runs reinforcement learning over both at once, so the model discovers what makes a better scaffold instead of engineers guessing.

An analogy: one worker is handed a hammer, another designs their own hammer and then swings it. The second is likely to end up with a more comfortable tool. In the team's own framing, the model improves its execution framework on its own, so the model is tuning the way code gets written and not just the code.

If that holds, the effects go beyond benchmark numbers. The evolution of coding tools would no longer be capped by what human engineers can imagine as a good scaffold, and the model could try execution strategies and tool combinations nobody designed. Once such a reinforcement loop is running, iteration could outpace manual tuning.

Pressure on closed-source coding agents

Coding agents have long been seen as closed-source models' last stronghold. The task is hard: it needs context within a complex repository, multi-step planning, many tool calls, handling error feedback and repeated revision, a different level of difficulty from writing a paragraph. That's why the biggest labs have put heavy bets here. Claude Code is said to be running at an annualized revenue above $2.5 billion, GitHub Copilot has tens of millions of users, and AI programming is arguably the fastest, most dependable revenue line in the business. The calculation behind it is that foundation models can be given away, while working coding agents are what you pay for.

Then a family appears with an MIT license, a GGUF build, local deployment, every size from small to huge and top open-source scores on six benchmarks. The team behind it is reportedly not a thousand-person organization. It's described as a research group doing post-training on existing open models, with no funding announcement and no keynote, just models, scores and a permissive license released into the community. If that's accurate, it's an awkward contrast with moats built at a cost of billions.

Open versus closed

Today's picture, roughly. On the open side stand Ornith-1.0, DeepSeek V4, known for strong coding, and Qwen 3.5, which made a notable jump in code ability, all under MIT or similar terms, all locally deployable and commercially usable for free. On the closed side are Claude Code, whose SWE-Bench result is estimated around 75 to 80 without independent public verification, GPT-5.5 with Codex, whose scores were never disclosed, and Gemini with Code Assist.

One detail deserves attention. Most closed-source SWE-Bench numbers lack public, independently reproducible data, while Ornith's are released openly and anyone can run the model to check them. What happens when an open model is more transparent, scores higher and costs nothing? Other tracks offer precedent. After LLaMA was open-sourced, closed foundation models lost bargaining power. After Stable Diffusion, Midjourney had to move from selling models to selling an experience. Agentic coding may follow the same script, and with sharper tools on the open side, since the heart of agentic coding is engineering as much as model strength, and engineering is where open communities are strongest. Ten thousand developers building on Ornith could produce ten thousand frameworks and vertical solutions, and no closed team is that large. The contest shifts from whose model is stronger to whose ecosystem is bigger.

Three signals

The first is that the model layer of agentic coding is turning into a commodity. When a 9B model reaches agent-level coding, a permissive license allows free commercial use and a GGUF build runs on a MacBook, the model stops being scarce, much as cloud computing made physical servers less scarce. Value moves toward scenario fit and workflow integration. Whoever wires Ornith into your repository, connects it to your CI/CD and adapts it to your coding standards and business logic earns the money, and the model itself costs nothing.

The second is that reinforcement learning that jointly optimizes scaffold and solution may become a new paradigm. Ornith isn't the first model post-trained with RL, but it pushes joint optimization of the two further than earlier work and backs it with six benchmark scores. The future described is one where models design their own frameworks, run them, receive RL feedback and iterate on the framework, a loop that could speed up quickly once closed.

The third is that open source has a structural edge in the agent era for the first time. Open models proved themselves in chat, writing and translation, but complex task execution is the first dimension where they've raced ahead. That's not accidental. Tool calling, multi-step planning and environment interaction are engineering problems, and engineering suits open collaboration. A closed company can hide weights, but it can't hide the best engineering architecture, and a thousand developers refining one agent framework will tend to outdo any single team's design.

An old question, again

When analysts at Goldman Sachs looked at the AI industry, they asked whether cheaper intelligence creates more demand or destroys pricing power. That question was about foundation models. It now applies to agentic coding. If a fully open, full-range, locally deployable family with the strongest open-source results is available to everyone under MIT, what happens to the economics of AI-written code? The Ornith team answered with its actions: release it first, and sort out the rest afterward.