The model invents its own tasks, builds the rig to test them, then trains on the results — and DeepReinforce gave the weights away

DeepReinforce released Ornith-1.5 on 19 August in three open-weight sizes: a 397-billion-parameter mixture of experts, a 35-billion one that activates 3 billion parameters per token, and a 9-billion dense model, all under an MIT licence with the weights on Hugging Face. The sizes are not the interesting part. Instead of learning from a fixed set of human-written tasks, the model proposes its own, scores each one for validity, difficulty and novelty, writes the scaffolding needed to run it, then generates attempts that feed back into reinforcement learning. The team calls this closing the loop from self-scaffolding to self-improvement. On the lab's own figures the 397B model scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, against 85.0 and 59.0 for Claude Opus 4.8 and 82.7 and 46.2 for GLM-5.2, plus 92.8 on GPQA Diamond. Its context window is 262,144 tokens and the full model is roughly 800 gigabytes in BF16, so open here does not mean it runs on your laptop. The 9B version, at 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-bench Verified, is the one that does.

✓ Verified · 4 sources

Related stories