122B MoE (approximately 125B including multimodal components)
262,144 tokens natively (approximately 1M with validated YaRN scaling)
MIT
MIT-licensed addition to the Ornith-1.5 family, released two months after the original 9B, 35B-A3B, and 397B checkpoints. Ornith-1.5-122B is a native image-and-video-capable reasoning and agent model built with the Qwen3.5 MoE conditional-generation architecture. Its configuration contains 48 language-model layers with a 3,072-dimensional hidden state, 32 query heads, 2 key-value heads, 256 routed experts with 8 selected per token, one Multi-Token Prediction layer, and a repeating pattern of three linear-attention layers followed by one full-attention layer. A 27-layer vision encoder supports image and video token inputs. The checkpoint provides a native 262,144-token context window, with the team validating static YaRN scaling by a factor of four for approximately one million tokens while warning that static scaling can slightly reduce quality on ordinary-length requests. It is an always-thinking model that emits a think block before its final response and produces structured tool-call blocks that vLLM and SGLang can expose through OpenAI-compatible reasoning_content and tool_calls fields. The checkpoint extends Ornith-1.5's self-improvement approach, in which training jointly optimizes generated tasks, task-specific scaffolds and evaluation harnesses, and solution rollouts through reinforcement learning. Ornith reports 81.2 on Terminal-Bench 2.1, 80.0 on CyberGym, 51.2 on NL2Repo, 76.9 on SWE-bench Pro, 84.5 on SWE-bench Verified, 76.4 on SWE-bench Multilingual, 48.0 on HLE with tools, 53.8 on SciCode, 78.5 on BrowseComp, 74.6 on MCP-Atlas, 70.8 on Toolathlon-Verified, and 79.4 on ClawEval. Software-engineering evaluations include network and package restrictions and explicit anti-reward-hacking checks. The full BF16 checkpoint occupies approximately 250.2 GB across 62 safetensors shards. Official deployment variants include FP8, NVFP4, GGUF in BF16 through Q4/Q5/Q6/Q8 formats with a separate vision projector, and MLX BF16, 4-bit, and 8-bit releases. The team's single-80GB-GPU deployment claim therefore refers to compressed variants rather than the BF16 checkpoint. Supported serving paths include Transformers 5.8.1 or newer, vLLM 0.19.1 or newer, SGLang 0.5.9 or newer, llama.cpp, Ollama, Hermes Agent, OpenClaw, OpenCode, and Unsloth.