DeepReinforce, the research lab that introduced the idea of a model writing its own coding scaffolding, has now taught that model to write its own homework too. Ornith-1.0, released in June, could build the tools, instructions, and orchestration logic around a task on its own, but a human still had to supply the task itself. Ornith-1.5, published Wednesday, removes that step too, extending the same self-scaffolding framework the earlier model introduced into a loop that also decides what to practice on.

Each training cycle now runs through three linked stages instead of one. The model looks at a codebase, a general instruction about the kind of problem to generate, and its own record of what it has already solved, then proposes a task that sits just past its current ability. It then builds a scaffold, the tool access and step-by-step decomposition a solver would need, and produces an attempt conditioned on both the task and the scaffold. One reward signal flows backward through all three stages, and a single training update improves the task generator, the scaffold builder, and the solver at once.

Reward for each generated task multiplies three separate checks. A validity gate confirms that a proposed task, paired with its scaffold, actually runs as a working, gradable environment, and zeroes out anything malformed. A difficulty score rewards problems that sit near the model’s current ceiling, tuned so a task pays out while the model still fails it roughly four times in five and stops paying once it gets solved too easily. A novelty check discounts tasks that resemble ones already generated. All three stages train with GRPO, a reinforcement learning method already common in open-model post-training.

On DeepReinforce’s own published tables, averaged across five runs, the 397B flagship posts 85.1 on Terminal-Bench 2.1. Its DeepSWE result lands at 56.0, a pair DeepReinforce says sits close to Claude Opus 4.8’s own 85.0 and 59.0. The company also claims a lead over GLM-5.2 and DeepSeek-V4-Flash-0731 at comparable parameter counts. The 35B model, which activates only 3B parameters per token, reaches 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified. The 9B dense model scores 47.0 and 70.6 on that same pair, which DeepReinforce says clears Gemma 4-31B and Qwen 3.6-35B. At flagship scale the model also clears 92.8 on GPQA Diamond. BrowseComp comes in at 86.6, pushing the story past coding into general reasoning and browsing agent work.

Every one of those numbers comes from DeepReinforce’s own harness. The release does not include benchmark results run by an independent lab, so the Opus and GLM comparisons should be read as self-reported until someone reproduces them outside DeepReinforce’s environment.

The benchmark table is not the interesting part of this release. The interesting part is a training loop that grades its own homework: the same system decides which problems are worth practicing, builds the environment those problems run in, and generates the transcripts used to reward itself. That setup is only as trustworthy as the validity gate sitting between the generator and the reward, because a model playing both teacher and student has every incentive to find whatever the gate will pass rather than what actually builds a stronger coder. DeepReinforce’s earlier open research on CUDA-L1 and its IterX agent loop ran into reward hacking repeatedly, which is presumably why the validity gate exists inside task generation now and not just inside solving. Whether it holds will show up in how these models perform on evaluations nobody at DeepReinforce wrote.

Ornith-1.0 shipped in June as four checkpoints, from 9B and 31B dense models up to 35B and 397B mixture-of-experts versions, post-trained on top of Qwen 3.5 and Gemma 4 and shipped under a permissive MIT license. That release treated the scaffold as something the model could learn rather than something an engineer wrote by hand. Ornith-1.5 pushes the same idea one step further back in the pipeline: instead of only learning to write scaffolding for a task, it now learns to write the tasks themselves. Weights for all three new sizes are posted to Hugging Face, and the 9B checkpoint also ships as a quantized build, branded Mobile, that runs directly on iPhone and Android.

None of this is independently verified yet. The collection only went up on Hugging Face this week, and no outside party has reproduced the closed-loop training run itself. Teams evaluating open coding models for internal agent stacks should treat the Terminal-Bench and SWE-Bench Verified scores as a starting point, not a final answer, and rerun the 35B checkpoint against their own repositories before committing a 2026 roadmap to a model that invented its own training data.

TestingCatalog (Nero Soares) reported the Ornith-1.5 release on August 19, 2026.