Ornith 1.0 is a new family of open-source models designed for agentic coding, released by deepreinforce-ai on HuggingFace. The family includes four model sizes — from a lightweight 9B dense model to a 397B MoE model — all post-trained using reinforcement learning (RL) on top of base models like Qwen 3.5 and Gemma 4.

The Family at a Glance

The Ornith 1.0 family consists of:

• Ornith-1.0-9B — Dense, ~9B parameters, suitable for single-GPU deployment (~19 GB in bf16)
• Ornith-1.0-31B — Dense, ~31B parameters
• Ornith-1.0-35B — Mixture of Experts (MoE), ~35B parameters
• Ornith-1.0-397B — MoE, ~397B parameters (largest member)

All models are MIT-licensed and available on HuggingFace under the deepreinforce-ai/Ornith-1.0-XXB namespace.

Training Approach: Self-Improving via RL

Ornith 1.0 uses a reinforcement learning training framework that goes beyond standard supervised fine-tuning. The model is trained to generate not only code solution rollouts but also the scaffold — the planning structure (like folder organization, file structure, test strategy) that drives those rollouts. By jointly optimizing the scaffold and the resulting solution, the model discovers better search trajectories and generates higher-quality solutions.

This "self-improving" approach is notable because it trains the model to think about code architecture before writing code, rather than jumping straight to implementation.

Benchmarks (Ornith-1.0-9B)

According to the official model card on HuggingFace, the 9B model achieves the following results on coding benchmarks:

Benchmark	Ornith-1.0-9B	Qwen3.5-9B	Qwen3.5-35B	Gemma4-12B	Gemma4-31B
Terminal-Bench 2.1 (Terminus-2)	43.1	21.3	41.4	21.0	42.1
Terminal-Bench 2.1 (Claude Code)	40.6	18.9	38.9	—	—
SWE-bench Verified	69.4	53.2	70.0	44.2	52.0
SWE-bench Pro	42.9	31.3	44.6	27.6	35.7
SWE-bench Multilingual	52.0	39.7	60.3	32.5	51.7
NL2Repo	27.2	16.2	20.5	10.3	15.5
Claw-eval Avg	63.1	53.2	65.4	32.5	48.5
SWE Atlas - QnA	17.9	9.2	13.2	—	—

Source: Ornith 1.0 README on HuggingFace (deepreinforce-ai/Ornith-1.0-9B). Results averaged over 5 runs. Some benchmarks have caveats in methodology notes.

Key Observations

• Strong 9B model: On Terminal-Bench 2.1 (Terminus-2), the 9B model (43.1) outperforms Qwen3.5-35B (41.4) and Gemma4-31B (42.1), demonstrating that RL-based training can significantly boost smaller models.
• SWE-bench Verified near-parity with 35B: The 9B model scores 69.4 on SWE-bench Verified, close to Qwen3.5-35B's 70.0. This is a strong result for a model of this size.
• Reasoning model: Ornith-1.0-9B is a reasoning model — it opens assistant turns with a <think> block before producing the final answer, enabling chain-of-thought reasoning.
• Tool use: The model emits well-formed function calls and supports OpenAI-style tool calling with vLLM and SGLang.
• MIT license: The models are globally accessible with no regional limitations.

Deployment

Ornith-1.0-9B can be served with vLLM, SGLang, or the HuggingFace Transformers library. It requires recent runtimes (transformers ≥ 5.8.1, vLLM ≥ 0.19.1, SGLang ≥ 0.5.9) and supports OpenAI-compatible API endpoints.

Limitations and Caveats

• Training data not disclosed: The exact training data and RL reward functions are not publicly documented.
• Benchmark methodology notes: The Terminal-Bench results include specific evaluation setup details (Harbor/Terminus-2 framework, JSON parser, temperature=1.0, 128K context, 4-hour timeout with 32 CPU cores and 48GB RAM). Reproducing results exactly may require matching these conditions.
• Smaller models may not generalize: The 9B model's strong results are impressive, but smaller models may struggle with complex multi-file projects or domains outside their training distribution.
• No community evaluation: Independent third-party evaluations of the full family are not yet available.

Conclusion

Ornith 1.0 represents a notable advance in open-source agentic coding models, particularly for the 9B size class. The RL-based training approach that optimizes both scaffold and solution is a promising direction, and the 9B model's ability to rival larger models on several benchmarks is noteworthy. The MIT license and open availability make it a practical choice for developers and researchers working with code-generation agents.