github.com2026-06-29authorityHN front pageabout the Ornith 1.0 familyHN 262 pts · 53 comments2,716 words
Ornith-1.5 is the latest release of a family of self-improving open-source models designed for agentic tasks. Both generations are MIT licensed and available as dense 9B models plus two Mixture-of-Experts models (35B, 397B). The 397B and 35B-A3B variants achieve 86.1 and 85.2 on Terminal-Bench 2.1 respectively, while the 9B model scores 46.2. All models support a 262,144-token context window and require runtimes like vLLM ≥ 0.19.1 and SGLang ≥ 0.5.9.written, reviewed, cited
- DeepSWE scores for Ornith-1.5-35B-A3B and Ornith-1.0-35B-A3B are exactly 0, indicating a complete failure on that specific benchmark.
- The DeepSWE evaluation harness explicitly disables network access to prevent the model from accessing prior solutions or commits.
- Ornith models output reasoning in a separate reasoning_content field by default, requiring a specific parser to surface the chain-of-thought.
deep-reinforce.com2026-06-25about the Ornith 1.0 familyHN 86 pts · 10 comments1,487 words
Ornith-1.0 is a self-improving family of open-source models designed for agentic coding tasks. The 397B MoE variant achieves 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified by jointly optimizing solution rollouts and task-specific harnesses. To prevent reward hacking, the system employs a frozen LLM judge that acts as a veto on top of the verifier rather than the primary reward.written, reviewed, cited
- The 9B Dense variant matches or exceeds performance of larger models like Gemma 4-31B and Qwen 3.6 35B.
- Ornith-1.0-35B surpasses Qwen 3.5-397B on Terminal-Bench 2.1 with a score of 64.4 versus 53.5.
- Evaluation of Terminal-Bench 2.1 uses a 128K context window with a 4-hour timeout and 32 CPU cores.
youtube.com2026-06-28about the Ornith 1.0 familyHN 1 pts · 0 comments
