Open-Source Ornith-1.5 Drops 397B MoE Model Under MIT License
The open-source Ornith-1.5 model family released 9B Dense, 35B MoE, and 397B MoE checkpoints trained with self-improving strategies. The top 397B MoE model claims performance comparable to Claude Opus 4.8 on coding benchmarks.
Impact: Medium
Why it matters
You can download, fine-tune, or host a top-tier open-weights MoE model for local and self-hosted AI coding workflows.
TL;DR
- 01Ornith-1.5 provides open-weights MoE checkpoints under an unrestricted MIT license.
- 02Utilize GGUF or MLX quantizations for local testing on workstation hardware.
- 03Evaluate the 397B MoE model as a self-hosted alternative for coding agent pipelines.
Key facts
- SWE-Bench Verified Score
- 86.0% (self-reported)
- SWE-Bench Pro Score
- 65.1% (self-reported)
- License
- MIT License
- Supported Formats
- FP8, GGUF, MLX, NVFP4
Self-Improving Agentic Architecture
Ornith-1.5 introduces an automated self-improvement loop for training foundation models. The system generates task proposals, creates custom execution scaffolds, and runs solution rollouts for reinforcement learning (RL):
- Model Variants: 9B Dense, 35B MoE, and 397B MoE.
- License: Permissive MIT License for unrestricted commercial and research use.
- Quantizations: FP8, GGUF, MLX, and NVFP4 formats published on Hugging Face.
SWE-Bench Performance Metrics
The flagship 397B MoE model achieves high scores across coding and software engineering benchmarks (self-reported by vendor):
- SWE-Bench Verified:
86.0% - SWE-Bench Pro:
65.1% - SWE-Bench Multilingual:
79.6%
Try it in 2 minutes
huggingface-cli download ornith/Ornith-1.5-397B-MoE-GGUF --include '*.gguf'bash
✓ When to use
- When building private on-premise AI coding tools requiring MIT license compliance.
- When running offline agentic workflows where cloud API access is restricted.
✕ When NOT to use
- When limited by low-RAM inference environments unable to host large MoE parameters.
What to do today
- Download GGUF or FP8 quantizations from Hugging Face for local testing.
- Benchmark Ornith-1.5 397B MoE against your internal software engineering tasks.
Sources