Ornith-1.5-122B Released Under MIT: A Single-Node 122B MoE Reasoning Model Scoring 76.9 on SWE-bench Pro

Background and the Problem

Autonomous coding agents and software engineering tools have historically demanded a painful compromise: top-tier commercial models (such as Claude 3.7 Sonnet) offer high benchmark solve rates but present strict data exfiltration risks and ballooning API costs; conversely, mid-sized open-weight models have suffered steep drop-offs when dealing with multi-file dependency resolution and complex terminal debugging. Enterprise engineering organizations urgently require frontier-grade coding performance contained within a single on-premises GPU node (such as an 8-GPU DGX server). On October 20, 2026, the Ornith AI team answered this demand by open-sourcing Ornith-1.5-122B under the MIT license.

Architecture and How It Works

1. Single-Node MoE Topology:

• Features 122 billion total parameters structured in a fine-grained mixture-of-experts configuration, designed specifically to saturate the NVLink interconnect and memory bandwidth of a single DGX node;

1. Immediate Multi-Runtime Support:

• Ready day-one on vLLM 0.19.1+, SGLang 0.5.9+, and Transformers 5.8.1+;
• Simultaneously published in production-ready FP8, local GGUF formats, and Apple Silicon MLX weights, indexed directly on Ollama as ornith-1.5:122b;

1. Reinforced Agentic Loop Alignment:

• Trained with specialized reinforcement learning across CodeAct sandboxes and multi-turn terminal execution traces, suppressing unproductive retry loops and embedding Git-diff-based hypothesis generation.

Benchmarks and Measured Results

Evaluated within the standardized OpenHands agent harness against prominent open baselines:

• SWE-bench Software Engineering Confrontation:
• SWE-bench Verified: Ornith-1.5-122B registers 84.5 points (surpassing Qwen3.8-Flash-Next at 82.8);
• SWE-bench Pro: Reaches a dominant 76.9 points, thrashing Qwen3.8-Flash-Next (43.9 points) by an unprecedented 33-point margin;
• SWE-bench Multilingual: Delivers 76.4 points (vs. Qwen's 73.0);
• System and Terminal Metrics:
• Scores 81.2 on TerminalBench 2.1 and 80.0 on CyberGym;
• Attains 74.6 on the MCP-Atlas tool protocol benchmark, closing within striking distance of 284B flagship models.

Getting Started for Developers

• Model Card & Weights: Access directly on Hugging Face at https://huggingface.co/ornith-ai/Ornith-1.5-122B;
• Official Announcement: Reference @ornith_ status https://x.com/ornith_/status/2112553477457268799;
• Engineering Recommendation: Teams implementing internal code copilots or autonomous PR triage can run Ornith-1.5-122B via vLLM in FP8 on an on-prem DGX, connecting seamlessly to OpenHands or Cline for zero-API-cost private coding autonomy.