Ornith drops Ornith-1.0 coding models that build their own scaffolds

Ornith-1.0 is a new family of open-source models built specifically for agentic coding, spanning from a 9B Dense model for edge devices up to a massive 397B MoE frontier variant.

Agentic coding goes far beyond single-file autocomplete. It requires models to break down tasks, run terminal tools, debug errors, edit across multiple repositories, and keep the whole codebase coherent over time.

Built on top of pretrained Gemma 4 and Qwen 3.5, the lineup includes 9B Dense, 31B Dense, 35B MoE, and 397B MoE checkpoints.

Learning to build its own scaffold

Most reinforcement learning setups rely on rigid, hand-engineered scaffolds to prompt and direct the agent. Ornith flips this dynamic by treating the scaffold itself as a learnable object that evolves alongside the coding policy.

At every step of RL, the model first generates or refines a task-specific scaffold—shaping how it handles memory, tool orchestration, and error recovery. It then executes the coding rollout inside that custom workflow. Rewards from the rollout flow back to both steps, teaching the model to produce both cleaner code and better orchestration harnesses.

Giving a model control over its own execution harness naturally invites reward hacking, such as peeking at test scripts or hardcoding expected return values. Ornith guards against this with three layers: an immutable outer execution environment, a deterministic monitor that zeros out rewards for out-of-bounds tool calls, and a frozen LLM acting as a veto judge.

Benchmark scores and hardware specs

On the benchmarks, the numbers are punchy. Flagship Ornith-1.0-397B scores 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-bench Verified, outpacing Claude Opus 4.7 (70.3 and 80.8) and topping rivals like Minimax M3 and DeepSeek-V4-Pro.

The smaller variants hit well above their weight class. Ornith-1.0-35B puts up 64.4 on Terminal-Bench 2.1, beating the much larger Qwen 3.5-397B (53.5) while outperforming Qwen 3.6-35B and Gemma 31B. Even the 9B model reached 69.4 on SWE-bench Verified, rivaling Gemma 4-31B.

Running the full 397B model takes serious iron: the uncompressed model weighs around 794GB, with an FP8 build trimmed to 400GB. However, the 35B checkpoint comes in at 70GB, and the GGUF Q4 quantized version sits at just 21GB, meaning you can comfortably run it on a single high-end consumer GPU.

All weights and quantization files are already up on Hugging Face for anyone looking to build local agents.

Why it matters

If you are building autonomous coding workflows or trying to host an internal coding copilot without sending sensitive intellectual property to closed APIs, Ornith-1.0 is definitely worth a spin today. The 9B and 35B models punch remarkably hard for their footprint, lowering the hardware ceiling for real multi-file coding agents.

We think the self-scaffolding approach is a clever blueprint for training agents: optimizing the orchestration loop alongside the model yields huge efficiency gains over brute-force scale. That said, heavily benchmark-optimized RL can sometimes stumble on messy, non-standard enterprise codebases, so you will want to test it against your own real-world repositories before swapping out your existing stack.