DeepReinforce

DeepReinforce is an AI research lab that builds the Ornith family of open-weight language models. It is best known for releasing frontier-class models under the plain MIT license at a time when most large open-weight releases carry custom terms with revenue or user thresholds attached.

The lab's distinguishing technique is self-scaffolding reinforcement learning. Rather than training against a fixed set of human-written task harnesses, an Ornith model proposes its own tasks, builds a scaffold for approaching each one, and generates the solution rollouts that become the reinforcement-learning signal. Rewards are optimized across both the strategy and the solution together, so task-specific approaches emerge without an engineer designing them by hand. The lab was previously known for CUDA-L1 and for the IterX code-agent optimization loop.

Ornith 1.0 shipped in June 2026. Ornith-1.5 followed on August 19, 2026 in three sizes: a 397 billion parameter mixture-of-experts flagship, a 35 billion parameter mixture-of-experts model that activates roughly 3 billion parameters per token, and a 9 billion parameter dense model with a quantized build for phones. The flagship scores 86.1 on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0, and 56.0 on DeepSWE against 59.0, which puts a fully permissive open-weight model within reach of a closed frontier flagship on agentic coding work.

The lab publishes no hosted API and no price list, so using Ornith means self-hosting. The flagship is roughly 800 gigabytes at bf16 precision and needs eight-way tensor parallelism, on the order of eight H200 cards, which puts it out of practical reach for individuals even though nothing in the license restricts them. The smaller sizes are the accessible ones, and the 9 billion parameter dense model is the only Ornith release most people will run locally.