Ornith-1.5-35B-A3B is a historic milestone for open-source AI: the first mid-sized model to truly rival frontier proprietary systems like Claude Opus 4.8 in reasoning, coding, and agentic workflows. With its revolutionary end-to-end self-improvement training loop achieving 86.1 on Terminal-Bench 2.1 and 86 on SWE-Bench verified, this MoE model demonstrates that open-source intelligence has caught up with — and in some metrics surpassed — the industry giants while maintaining unrestricted commercial licensing.
On August 20, 2026, the landscape of open-source artificial intelligence shifted fundamentally. Ornith has officially released the Ornith-1.5 model family, a collection of LLMs that represents more than just a marginal improvement in parameter count. This release marks a historical milestone where open-source models are no longer just chasing proprietary giants—they are setting new benchmarks for autonomous intelligence.
At the heart of this release is the Ornith-1.5-35B-A3B, a Mixture-of-Experts (MoE) model that defies the traditional scaling laws. While many models struggle to maintain reasoning capabilities as they optimize for efficiency, Ornith-1.5 utilizes a unique self-improvement loop that allows it to bridge the gap between mid-sized models and massive frontier systems.
The Ornith-1.5 family is designed with scalability in mind, offering a range of architectures to suit different deployment needs. The family includes a 9B Dense model for edge applications, the flagship 35B MoE (A3B) for high-performance reasoning, and a massive 397B MoE variant for ultra-complex enterprise workloads.
What truly sets Ornith-1.5 apart is its training methodology. Building upon the self-scaffolding strategies introduced in Ornith-1.0, the 1.5 series implements a complete self-improvement loop. During training, the model does not just consume static datasets; it proposes its own new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning.
| Model Variant | Architecture Type | Best Use Case |
|---|---|---|
| 9B Dense | Standard Transformer | Edge deployment, low-latency needs |
| 35B MoE (A3B) | Mixture-of-Experts | High-performance reasoning, agents |
| 397B MoE | Mixture-of-Experts | Enterprise workloads, complex RAG |
The performance metrics for the Ornith-1.5-35B-A3B are nothing short of staggering. In coding and software engineering tasks, the model demonstrates a level of autonomy that was previously thought to be the exclusive domain of much larger models.
On the SWE-Bench, it achieved a verified score of 86, proving its ability to handle real-world repository-level engineering tasks with high precision. Beyond coding, the model excels in complex reasoning and tool usage.
💡 The Benchmark Breakthrough: Ornith-1.5-35B-A3B performs on par with Claude Opus 4.8 across reasoning, agentic workflows, and complex instruction following while using fewer than half the parameters of typical frontier models.
| Benchmark | Ornith Score | Significance |
|---|---|---|
| Terminal-Bench 2.1 | 86.1 | State-of-the-art terminal automation |
| SWE-Bench Verified | 86 | Real-world repository-level engineering |
| SWE-Bench Pro | 65.1 | Professional coding standards |
| SWE-Bench Multilingual | 79.6 | International development workflows |
| DeepSWE | 56 | Deep software understanding |
| HLE (Humanity's Last Exam) | 44.6 | Complex multi-step reasoning |
| ClawEval | 81.4 | General agentic capability |
| Tool Decathlon | 71.2 | Tool use proficiency |
The most revolutionary aspect of Ornith-1.5 is its training methodology. Unlike traditional models that consume static pre-training datasets, Ornith-1.5 implements a complete self-improvement loop:
This creates a continuous cycle of learning where the model effectively becomes its own teacher, constantly identifying areas for improvement and generating the exact data needed to address those weaknesses.
Because Ornith-1.5 is released under the MIT License, developers have ultimate flexibility: you can host it yourself using quantized versions or utilize managed API endpoints. This dual approach allows for both cost-sensitive local development and scalable production deployments.
The Ornith-1.5-35B-A3B is purpose-built for high-agency environments. Its strength in "Tool Decathlon" (71.2) makes it an ideal engine for autonomous agents that must interact with operating systems, APIs, and complex software environments.
| Use Case | Key Metric |
|---|---|
| Autonomous Coding Agents | SWE-Bench: 86 |
| System Administration & Terminal Automation | Terminal-Bench: 86.1 |
| Complex Reasoning & Logical Deduction | HLE: 44.6 |
| Multilingual Enterprise RAG | SWE-Bench ML: 79.6 |
Getting started with Ornith-1.5 is seamless. Developers can pull the weights directly from the official Ornith repository or use popular frameworks like Hugging Face and vLLM for deployment. For those working on Apple Silicon, the MLX-quantized versions are ready for immediate use, providing high-performance local inference.
If you are looking to integrate the model into a production pipeline, we recommend starting with the 35B MoE variant to balance cost and capability. For edge-case testing or low-latency requirements, the 9B Dense model serves as an excellent entry point into the Ornith ecosystem.
⚡ Pro Tip: The 35B MoE variant's "Active 3B" architecture means only 3 billion parameters are activated per token, ensuring lightning-fast inference speeds while maintaining frontier-level intelligence.
This article is based on official Ornith-1.5 release information and technical documentation. For the most current information and weights, please refer to the official sources:
🤖 Featured AI Developer Tools