Open Source AI Model Outperforms Claude Opus on Coding Benchmarks

Ornith AI Model Achieves Superior Coding Benchmark Performance

An open-source AI model, referred to as Ornith, has demonstrated superior performance on two key coding benchmarks, Terminal-Bench and SWE-bench Verified, surpassing Claude Opus 4.7.

Benchmark Performance:

• Ornith's Flagship Model:
  • Terminal-Bench: 77.5
  • SWE-bench Verified: 82.4
• Claude Opus 4.7:
  • Terminal-Bench: 70.3
  • SWE-bench Verified: 80.8

Ornith also outperforms other models like MiniMax and DeepSeek at its size.

Training Methodology:

• Self-Scaffolding: The model does not rely on human-written harnesses. Instead, it generates its own scaffolding at each step.
• Reward Training: A reward mechanism trains both the orchestration and the answer generation processes.
• Anti-Cheating Measures: To prevent the model from 'cheating' during training:
  • The environment is locked.
  • The team monitors for any 'peeking'.
  • A frozen judge is run on top to evaluate outputs.

Accessibility and Hardware Requirements:

• 397B Flagship Model: Requires data-center hardware for operation.
• 9B Model:
  • Size: 5GB
  • SWE-bench Verified Score: 69.4
  • Can run on a laptop.
• Licensing: Available under the MIT license for all sizes, with no additional restrictions.

Community Discussion:

The post, shared by @aq_mern_stack_dev on Instagram, prompts a discussion on whether users would switch their daily AI driver to an open model or continue using closed models.

Key Hashtags: #AI, #OpenSource, #Ollama

Related Comments:

• A user points out that the benchmark scores cited are for the 397B model, not necessarily the smaller ones.
• Another user reports that the 9B model struggled with a simple Perl script, producing syntax errors and logical flaws when reviewed by GPT.
• A link to ornith.site was shared.
• Questions were raised about the specific configuration required to run the 9B model on a laptop and its Tokens Per Minute (TPM).
• Confirmation was sought if the reported performance was from the 9B model.
• A user from Russia shared their experience, stating that while the 9B model is better than Qwen 3.5's 9B, it's not superior to Qwen 3.6's 9B. The 35B version works well but is not considered extraordinary.
• One comment expresses skepticism, suggesting it might be a text-only model.

Post Details:

• Platform: Instagram Reel
• Owner: Abdul Qadeer (@aq_mern_stack_dev)
• Timestamp: 2026-06-29T08:30:39Z
• Video Duration: 82.733 seconds
• Likes: 364
• Video Views: 18053 (Play Count)
• URL: https://www.instagram.com/reel/DaKcPZkA8TH/