DeepReinforce has released Ornith 1.0, an open-weights agentic coding model that runs on a single GPU yet posts frontier-level scores on agent benchmarks. The 35B Mixture-of-Experts model activates only about 3B parameters per token, ships under a permissive MIT license, and landed on Hugging Face on June 25, 2026.

Try It: Run a Coding Agent Locally

Ornith 1.0 is designed to drop into the agent tools you already use. It serves an OpenAI-compatible endpoint through vLLM or SGLang, so you can point an existing framework at a local address instead of a metered cloud API. It works out of the box with OpenHands, Hermes Agent, and OpenCode, while the GGUF build runs in Ollama, llama.cpp, and LM Studio. For a builder, that means a capable code agent that reads, writes, and runs across a repo without per-token charges and without your codebase leaving your machine.

Why It Matters for Creators

Most strong coding agents still sit behind paid cloud APIs, which is exactly the local versus cloud tradeoff builders weigh every day. Ornith pushes a genuinely usable agent onto local hardware: it scores 64.2% on Terminal-Bench 2.1 and 75.6% on SWE-bench Verified, beating Qwen3.5-35B, Qwen3.6-35B, and Gemma4-31B at the same size and edging past models many times larger. If you build apps, websites, or automations with an AI assistant, a local model at this level changes the cost math and keeps private work private.

Key Details

Model: 35B Mixture-of-Experts, roughly 3B active parameters per token, built on a Qwen 3.5 base.

License: MIT, so commercial use is allowed.

Benchmarks: Terminal-Bench 2.1 64.2%, SWE-bench Verified 75.6%, SWE-bench Pro 50.4%.

Serving: OpenAI-compatible endpoint via vLLM 0.19.1 or later (or SGLang 0.5.9 or later), with native tool calling. DeepReinforce trained the model with a self-scaffolding reinforcement-learning method described in its technical writeup.

What to Do Next

Pull the GGUF into Ollama for a quick local test, then wire the OpenAI-compatible endpoint into your agent of choice and run it against a real branch before deciding whether to swap it in for routine coding work. If you currently lean on cloud coding assistants, where the tooling keeps consolidating, treat Ornith as a local option worth benchmarking against your daily driver.