Ornith 1.0 35B

DeepReinforce's open agentic-coding model — a 35B-parameter Mixture-of-Experts post-trained on top of Gemma 4 + Qwen 3.5 with a self-scaffolding RL method (the model learns to generate its own task harnesses during training). 262K context, <think> reasoning, and native tool-calling for agentic dev workflows. SWE-bench Verified 75.6 / Terminal-Bench 2.1 64.2 — state-of-the-art among similarly-sized open models.

Runs locally via llama.cpp/Ollama from GGUF quants; Q4_K_M is ~21.2 GB — fits a 24 GB card but tight on context, since an MoE keeps all experts resident in VRAM. For 8-16 GB cards use the 9B build.