GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Ornith-1.0 Writes Its Own RL Training Scaffolds: Open-Source 397B Matches Claude Opus 4.7 on SWE-Bench

DeepReinforce released Ornith-1.0 on June 25, a family of open-source coding models built around a reinforcement learning architecture that discards human-designed training harnesses. Where every prior RL-tuned coding model trains against a fixed scaffold written by engineers, Ornith trains the model to produce that scaffold itself. The model then conditions its solution on the scaffold it wrote. Reward from successful task completions flows back through both stages, so per-task strategies emerge from the training process without manual harness engineering.

The weights and technical report are published on Hugging Face under the MIT license, which removes the commercial-use restrictions present in several recent permissive open-weight releases.

Model Family

Four size points, two architectures:

  • 9B Dense — edge deployment target; runs on resource-constrained hardware
  • 31B Dense — mid-range; comparable deployment profile to Qwen 3.6-27B
  • 35B MoE — efficient inference at frontier-adjacent performance
  • 397B MoE — flagship; post-trained on Gemma 4 and Qwen 3.5 pretrained checkpoints

The Gemma 4 and Qwen 3.5 foundation choices mean teams running those models can benchmark directly against Ornith using their existing infrastructure.

Benchmark Performance

Flagship (397B MoE) numbers from DeepReinforce’s internal evaluations:

BenchmarkOrnith-1.0-397BClaude Opus 4.7DeepSeek-V4-ProMiniMax M3
SWE-bench Verified82.480.880.680.5
Terminal-Bench 2.1 (Terminus-2)77.570.364.064.0
Terminal-Bench 2.1 (Claude Code)78.269.766.5—
SWE-bench Pro62.264.355.459.0
SWE-bench Multilingual78.9—76.2—

The 397B model leads on SWE-bench Verified and puts up a notably larger lead on Terminal-Bench 2.1, where Ornith scores 7.2 points above Claude Opus 4.7 on the Terminus-2 harness. SWE-bench Pro, which uses contamination-resistant proprietary tasks, is tighter: Claude Opus 4.7 at 64.3 edges Ornith at 62.2.

For the 9B model, SWE-bench Verified at 69.4 and Terminal-Bench 2.1 at 43.1 outperforms comparably sized alternatives including Qwen 3.5-9B (53.2 SWE, 21.3 TB) and Gemma 4-12B (44.2 SWE, 21.0 TB).

The Self-Scaffold Mechanism

Standard RL training for coding agents uses a human-engineered harness: a fixed scaffold that defines how the model approaches a problem, breaks it into steps, calls tools, and evaluates partial solutions. Writing and tuning those harnesses is one of the primary costs of building capable coding agents. Ornith eliminates that dependency by treating scaffold generation as a learned capability rather than a design input.

Each RL training step runs in two stages. First, the model generates a task-specific scaffold. Then it generates a solution conditioned on that scaffold. Reward signal propagates back through both stages. Over many training steps, domain-specific problem-solving strategies emerge without anyone specifying what they should look like.

The practical implication is that the model can specialize its approach to a task class automatically. Whether that transfers cleanly to task distributions not represented in training is the open question for teams evaluating the model on their own workloads.

Context

Ornith-1.0 follows DeepReinforce’s earlier open-source work: CUDA-L1 (GPU kernel optimization) and the IterX code-agent optimization loop. The self-scaffold mechanism extends IterX’s approach to the training process itself rather than only inference-time iteration.

The numbers are self-reported and not yet independently replicated on the standard swebench.com leaderboard. The benchmark table includes GLM-5.2, which scores 81.0 on Terminal-Bench 2.1 and is absent from the SWE-bench Verified comparison, suggesting Ornith-1.0 may trail GLM-5.2 on that specific benchmark. DeepReinforce has not addressed that comparison directly.

MIT license and immediate availability put Ornith-1.0 ahead of most comparably capable open-weight releases on deployment friction. Teams building coding agents should run the 35B MoE variant against their specific task distribution before drawing conclusions from the top-line numbers.