Ornith-1.0 is DeepReinforce’s open-source, self-improving family of LLMs built specifically for agentic coding. It spans the full spectrum — from a compact 9B Dense model that runs on edge devices, through 31B Dense and 35B MoE, up to a 397B MoE frontier-scale flagship — all built on top of pretrained Gemma 4 and Qwen 3.5.
The headline isn’t just the scores, it’s how they were trained. Rather than rely on a human-designed harness to drive reinforcement learning, Ornith-1.0 learns to write its own scaffold alongside each solution — a “self-improving” loop DeepReinforce calls self-scaffolding. The result is state-of-the-art performance among open-weight models of comparable size across a broad range of agentic coding benchmarks.
At flagship scale, Ornith-1.0-397B matches Claude Opus 4.7 on Terminal-Bench 2.1 and SWE-Bench Verified and edges out leading open weights like DeepSeek V4-Pro and MiniMax M3. The honest caveat: it still trails the newest closed frontier (Claude Opus 4.8) on the hardest splits.
What is Ornith-1.0?
Ornith-1.0 is a model family, not a single checkpoint. The variants let you trade cost for capability:
- 9B Dense — edge-deployable; still matches much larger models like Gemma 4-31B.
- 31B Dense — a practical single-GPU-class workhorse.
- 35B MoE — strong price/performance; punches well above its weight.
- 397B MoE — frontier-scale flagship, competitive with the closed leaders.
Every variant targets the same job: long-horizon, tool-using coding agents. All ship as open weights on Hugging Face, so you can self-host the size that fits your hardware.
The self-scaffolding training framework
The core idea is a self-improving RL loop that jointly learns to solve tasks and to construct the scaffolds that guide those solutions. Most pipelines rely on a fixed, human-designed harness shared across a task category; Ornith-1.0 instead treats the scaffold as a learnable object that co-evolves with the policy.
- Two-stage RL step. Conditioned on a task and its previous scaffold, the model first proposes a refined scaffold; conditioned on that scaffold, it then generates a solution rollout.
- Shared reward. Reward from the rollout flows back to both stages, so the model is optimized not only to produce better answers but to author the orchestration that elicits them.
- Emergent strategies. Over training, scaffolds are continually mutated and selected toward those that induce higher-reward trajectories — per-task-category strategies emerge automatically, without hand-engineered harnesses.
Letting a model author its own scaffold invites reward hacking, so DeepReinforce adds three defenses: an immutable trust boundary (environment, tools and test isolation stay outside the model’s reach), a deterministic monitor that zero-rewards any attempt to read withheld paths or edit verification scripts, and a frozen LLM judge that acts as a veto on top of the verifier.
Benchmarks
Flagship Ornith-1.0-397B against leading models on agentic coding (higher is better; figures are DeepReinforce-reported as of June 2026 — treat them as a dated snapshot).
| Benchmark | Ornith-1.0-397B | Claude Opus 4.7 | DeepSeek-V4-Pro | Qwen3.5-397B |
|---|---|---|---|---|
| Terminal-Bench 2.1 (Terminus-2) | 77.5 | 70.3 | 64.0 | 53.5 |
| SWE-Bench Verified | 82.4 | 80.8 | 80.6 | 76.4 |
| SWE-Bench Pro | 62.2 | 64.3 | 55.4 | 51.6 |
A few things worth knowing:
- It leads open weights on Terminal-Bench. At 77.5 the 397B is well clear of DeepSeek V4-Pro and MiniMax M3 (both ~64) on the Terminus-2 harness.
- SWE-Bench Verified is a near-tie at the top. 82.4 edges Claude Opus 4.7 (80.8) and DeepSeek V4-Pro (80.6).
- SWE-Bench Pro is where closed still leads. On the hardest, leakage-controlled split it trails Claude Opus 4.7 (64.3) — the gap to the newest closed frontier is real.
The smaller variants are the surprise. Ornith-1.0-35B posts 64.2 on Terminal-Bench 2.1 and 75.6 on SWE-Bench Verified — even surpassing Qwen3.5-397B on Terminal-Bench (64.2 vs 53.5) at a fraction of the size. The edge-deployable 9B still reaches 43.1 / 69.4, matching much larger models such as Gemma 4-31B.
See the official announcement for full methodology (harnesses, temperatures, context windows and run averaging).
How Ornith-1.0 compares
The June 2026 open-weight coding field is crowded, but Ornith-1.0 carves out a clear niche — strongest agentic/terminal behaviour among open weights at the top end, plus a genuinely useful small-model tier.
- vs DeepSeek V4-Pro: roughly tied on SWE-Bench Verified, but the 397B pulls clearly ahead on Terminal-Bench (77.5 vs 64.0). DeepSeek’s MIT license and cheap hosted API are still a draw.
- vs MiniMax M3: Ornith leads on the terminal/agentic harness; both are open weight.
- vs Claude Opus 4.8: the closed frontier still wins the hardest splits (SWE-Bench Pro, Terminal-Bench), but Opus costs $5/$25 per 1M tokens while Ornith is free to self-host.
For a wider view of where open and closed models sit, see our 2026 LLM landscape guide.
How to access Ornith-1.0
- Self-host the open weights from Hugging Face (
deepreinforce-ai), picking the size that fits your hardware — the 9B/31B Dense or 35B MoE for modest setups, the 397B MoE only with serious multi-GPU infrastructure. - There is no official hosted API or pricing at launch — deployment is self-hosted.
- Review the model card’s license before commercial use, especially given the Gemma 4 / Qwen 3.5 lineage.
Best for
- Agentic coding and terminal / SWE-style software tasks
- Self-hosted coding assistants, including edge deployments with the 9B
- Research on self-improving RL and learned agent scaffolds
Pros & cons
Strengths
- Open weights across a wide range — 9B / 31B Dense and 35B / 397B MoE
- Flagship 397B rivals Claude Opus 4.7 on Terminal-Bench 2.1 and SWE-Bench Verified
- Edge-deployable 9B matches much larger models like Gemma 4-31B
- Novel self-improving "self-scaffolding" RL training (the model learns its own harness)
Limitations
- The 397B MoE needs serious multi-GPU infrastructure to self-host
- No official hosted API or pricing announced yet
- Trails the very latest closed frontier (e.g. Claude Opus 4.8) on the hardest benchmarks
- Built on Gemma 4 / Qwen 3.5 — base-model license terms may apply
Frequently asked questions
Is Ornith-1.0 open source?
Yes. Ornith-1.0 ships as open weights (released on Hugging Face) across multiple sizes. Review the model card's license before commercial use, as it builds on Gemma 4 and Qwen 3.5 whose base terms may apply.
What sizes does Ornith-1.0 come in?
A full spectrum — 9B Dense and 31B Dense, plus 35B MoE and a flagship 397B MoE — so you can pick edge-friendly or frontier-scale variants.
What is Ornith-1.0 best at?
Agentic coding. The 397B variant posts 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, matching or beating Claude Opus 4.7 on both.
What is "self-scaffolding" training?
Instead of a fixed human-designed harness, Ornith-1.0 learns to generate both the task-specific scaffold and the solution rollout, jointly optimizing them so better search strategies emerge automatically.
How small can it go for edge devices?
The 9B Dense variant is edge-deployable and still scores 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified, matching much larger models such as Gemma 4-31B.
Is there a hosted API?
At launch DeepReinforce released open weights without an official hosted API or pricing; you self-host the weights.
What hardware do I need to run the 397B?
As a 397B-parameter MoE it needs substantial multi-GPU infrastructure. For lighter setups, the 35B MoE or 9B/31B Dense variants are far more practical.
How does it compare to closed models?
The flagship 397B is competitive with Claude Opus 4.7 on coding benchmarks while remaining open weight, though it trails the newest closed frontier on the hardest tasks.