Ornith-1.0: DeepReinforce and Its High-Performing Open LLMs

Ornith-1.0: A Promising Model from DeepReinforce

Ornith-1.0 marks a significant advancement in the field of large language models (LLMs) with its open MIT license. It is the first release from DeepReinforce, a company that remains somewhat discreet about its activities.

The model comes in several variants: 9B Dense, 31B Dense, 35B MoE, and 397B MoE. These variants are based on the pretrained models Gemma 4 and Qwen 3.5, both licensed under Apache 2.0. This licensing compatibility allows for flexible use without additional constraints, unlike some earlier models from Gemma.

Ornith-1.0 achieves state-of-the-art performance among comparable open-source models on coding benchmarks, making it a powerful tool for developers.

Performance and Testing

Ornith-1.0 was tested with the file ornith-1.0-35b-Q4_K_M.gguf (20 GB) via LM Studio, connected to Pi. The initial results are promising, with the model effectively executing agent frameworks across various tool calls.

During a session, the model was able to locate the code to decode an actor cookie and open an insertion dialog on a Datasette checkout. These tasks were performed with ease, demonstrating the model's ability to handle complex operations.

Additionally, Ornith-1.0 drew a pelican at a speed of 103 tokens/second. Although the drawing was slightly distorted, the pelican remains recognizable.

DeepReinforce: A Discreet Company

Little information is available about DeepReinforce. The company's first notable paper, titled CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning, dates back to June 2025. This hints at a potential for ongoing innovation in the field of reinforcement learning.