Ornith-1.0: Open Source Model Beats Models 10X Its Size Even Opus 4.7

What Is Ornith-1.0?

Ornith-1.0 is a new open-source model family from a lab called DeepReinforce, named after the Greek word for bird. It's built for one job: agentic coding — going into a terminal, writing code, running it, fixing it when it breaks, and finishing the task, rather than just chatting or writing essays.

The models are built on top of Gemma 4 and Qwen 3.5, and DeepReinforce claims state-of-the-art results among open models of similar size on coding benchmarks.

The Model Family: Four Sizes, One Architecture

Ornith-1.0 isn't a single model — it spans a huge range, from edge-friendly to frontier-scale:

The Core Innovation: Self-Written Scaffolding

When you train an AI agent with reinforcement learning to do coding tasks, you usually need a harness — scaffolding around the model that tells it how to explore the problem, what tools to call, and how to structure its attempts. Normally, humans design that harness by hand for each category of task.

Ornith does it differently: the model learns to write its own scaffolding, then uses that scaffolding to generate its solution. Every training step, it does two things back to back — first it looks at the task and the scaffold it used last time and proposes an improved version, then it attempts the task using that new scaffold.

The reward that comes back updates boththe scaffold-writing and the solution-writing. Over thousands of steps, the scaffolds evolve on their own, selected for whichever versions produce the best results — the model isn't just getting better at coding, it's getting better at designing its own process for coding.

Benchmark Results: How Ornith Stacks Up

At the top end, Ornith-1.0-397B scores 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified — ahead of Claude Opus 4.7 (70.3 and 80.8) and other open-weight models in its class, including MiniMax M3 and DeepSeek-V4-Pro. To be fair, Claude Opus 4.8still comes out ahead of Ornith on most of these benchmarks — the claim is "best among open-source models," specifically beating the previous Opus generation.

The 35B model is the most eyebrow-raising result: it beats Qwen 3.5's 397B parameter model on Terminal-Bench — a model more than ten times bigger, losing to the 35B version. At the small end, the 9B model lands in the same range as models several times its size, like Gemma 4-31B. The pattern across all three size classes is the same: smaller footprint, competitive or better results.

Handling Long Agentic Training at Scale

Ornith-1.0 uses an asynchronous, pipelined RL setup to deal with long agentic training runs where the policy and the data used to train it can drift apart over time. They handle this with a staleness weight that gradually discounts older, more out-of-date tokens and drops them completely past a certain age threshold.