DeepReinforce has released Ornith-1.0, an open-source model family built for agentic coding. The lineup spans four sizes, from a 9B dense model to a 397B mixture-of-experts flagship. Every checkpoint ships under the MIT license on Hugging Face. The models are post-trained on top of pretrained Gemma 4 and Qwen 3.5.

Most coding agents pair a model with a fixed, human-designed harness. Ornith-1.0 instead learns to write its own. The DeepReinforce research team reports state-of-the-art results among open models of comparable size.

TL;DR

• Ornith-1.0 ships in 9B, 31B, 35B-MoE, and 397B-MoE sizes under MIT, built on Gemma 4 and Qwen 3.5.
• The model learns its own scaffold during RL, jointly optimizing the harness and the solution.
• Ornith-1.0-397B tops Claude Opus 4.7 on both headline benchmarks, but not Opus 4.8 or the larger GLM-5.2-744B.
• Three layers — fixed trust boundary, deterministic monitor, frozen LLM judge — guard against reward hacking.

What is Ornith-1.0?

Ornith-1.0 is a set of reasoning models tuned for coding agents. The variants are 9B Dense, 31B Dense, 35B MoE, and 397B MoE. The 35B model is mixture-of-experts and activates roughly 3B parameters per token. FP8 and GGUF builds are also published for faster local serving.

Each model is a reasoning model. Replies open with a <think> block before the final answer. The serving recipes enable a reasoning parser, so that trace returns in a separate reasoning_content field. The models also emit well-formed tool calls for agent loops.

Deployment is straightforward. The 9B model is about 19GB in bf16 and serves on a single 80GB GPU. Serving recipes target vLLM, SGLang, and Transformers. Each model exposes an OpenAI-compatible endpoint. Standard agent frameworks therefore work without code changes.

Interactive Explainer

The Self-Scaffolding Idea

Most coding agents rely on a scaffold, also called a harness. A scaffold wraps the model with memory, tools, error handling, and orchestration logic. AI teams usually hand-design one scaffold per task category.

Ornith-1.0 treats the scaffold as a learnable object instead. During reinforcement learning, the scaffold co-evolves with the model’s policy. Each RL step runs in two stages.

First, the model reads the task and its previous scaffold. It then proposes a refined scaffold. Second, it uses that scaffold and the task to generate a solution rollout. Reward from the rollout flows back to both stages.

So the model is optimized to author orchestration, not just answers. Over training, higher-reward scaffolds are mutated and selected automatically. Per-task strategies emerge without hand-engineered harness design.

Training also runs asynchronously, using a pipeline-RL setup. A staleness weight downweights older, off-policy tokens and drops them past a threshold. The optimization uses a token-level GRPO objective.

Guarding Against Reward Hacking

Letting a model write its own scaffold invites reward hacking. A scaffold could read visible test files and hardcode expected outputs. It could also copy an oracle solution sitting in the environment. DeepReinforce team describes three defense layers.

1. The outer trust boundary is fixed and immutable. The environment, tool surface, and test isolation stay outside the model’s reach. The model evolves only its inner policy scaffold.
2. A deterministic monitor flags banned actions. Reading withheld paths or editing verification scripts earns zero reward. Those trajectories are excluded from the advantage computation.
3. A frozen LLM judge acts as a veto. It sits on top of the verifier, not as the primary reward.

Benchmark

DeepReinforce reports vendor numbers across several agentic coding benchmarks. At flagship scale, Ornith-1.0-397B posts 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified. On SWE-Bench Verified, that 82.4 trails only Claude Opus 4.8 (87.6) among the listed models. On Terminal-Bench 2.1, the picture is more mixed.

Ornith-1.0-397B beats Claude Opus 4.7 (70.3) on Terminal-Bench 2.1. But it trails Claude Opus 4.8 (85) and the larger GLM-5.2-744B (81.0). So the ‘state-of-the-art’ claim is scoped to open models of comparable size.

The smaller models carry the efficiency case. The 35B model scores 64.2 on Terminal-Bench 2.1, above Qwen 3.5-397B’s 53.5. The 9B model reaches 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified.

Use Cases and a Quick Start

The models target terminal-native coding agents and repository-scale work. Practical fits include multi-file refactors, bug localization, and test-driven patches. The 9B model suits edge or single-GPU setups where latency and cost matter. The 397B model targets maximum accuracy on long, multi-step tasks.

For example, a dev can run the 9B model locally to triage a failing test suite. A platform team can self-host the 397B model for an internal coding agent.

Serving is a one-liner with vLLM:

vllm serve deepreinforce-ai/Ornith-1.0-9B \
    --served-model-name Ornith-1.0-9B \
    --max-model-len 262144 \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --trust-remote-code

Then call it with any OpenAI client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

resp = client.chat.completions.create(
    model="Ornith-1.0-9B",
    messages=[{"role": "user", "content": "Write a Python is_prime(n)."}],
    temperature=0.6, top_p=0.95,
)
msg = resp.choices[0].message
print(getattr(msg, "reasoning_content", None))  # the <think> trace
print(msg.content)                              # the final answer

The reasoning trace returns in reasoning_content, with the answer in content. Recommended sampling is temperature=0.6, top_p=0.95, top_k=20. The model also plugs into OpenHands, OpenClaw, and OpenCode.