Ornith releases Ornith-1.5-122B as MIT open weights

The card calls Ornith-1.5-122B a 122B mixture-of-experts model, the most capable and efficient member of the Ornith-1.5 family, and says it is designed for single-GPU deployment[1]. It gives no active-parameter count. The model page lists the checkpoint as 125B parameters in BF16. The safetensors metadata counts 125,086,497,008 parameters, about 250 GB at 16-bit, and the repository stores 250.2 GB[1][14].

The published config.json sets the architecture to Qwen3_5MoeForConditionalGeneration, with 256 experts, 8 experts per token and a shared expert. max_position_embeddings is 262,144, the same context window the card states[10][1]. That config also includes a vision encoder, and the GGUF repository includes a 0.9 GB vision-projector file, mmproj-Ornith-1.5-122B-BF16.gguf. The card's quickstart and benchmark notes cover text[10][4].

The citation on the card points at an August 2026 post, "Ornith-1.5: From Self-Scaffolding to Self-Improvement", credited to the Ornith Team. That post describes 397B, 35B-A3B and 9B models. It does not mention a 122B[1][13].

Alongside the BF16 checkpoint, Ornith published FP8 (about 133 GB of files), NVFP4 (about 78 GB), MLX (about 244 GB) and MLX builds at 8-bit (about 130 GB) and 4-bit (about 69 GB)[3][8][5][7][6]. The GGUF repository has five model files: BF16 at 249.4 GB, Q8_0 at 132.6 GB, Q6_K at 102.4 GB, Q5_K_M at 88.7 GB and Q4_K_M at 75.8 GB[4]. Ollama lists ornith-1.5:122b in a 73–77 GB range, with a 262,144-token context, and the run command is ollama run ornith-1.5:122b[9].

The card says this is a reasoning model: the assistant turn opens with a think block, and the recipes split that into a reasoning_content field and parse <tool_call> blocks into OpenAI-style tool calls. It asks for Transformers 5.8.1 or newer, vLLM 0.19.1 or newer, or SGLang 0.5.9 or newer. The vLLM and SGLang commands load ornith-ai/Ornith-1.5-122B at a 262,144-token context, using Qwen3 reasoning and tool-call parsers. The llama.cpp command loads the GGUF repository at the same context length. The card also shows how to point OpenCode, Hermes Agent and OpenClaw at that local server[1].

For a window past 262,144 tokens, the card suggests YaRN RoPE scaling. A factor of 4.0 is the setting it says reaches about 1 million tokens. It warns that current open-source runtimes apply one scale to every request, which can hurt quality on ordinary lengths, and that the factor should match the window you actually need. A factor of 2.0 is the example it gives when requests top out around 524,288 tokens[1].

For general tasks the card recommends temperature 1.0, top_p 0.95, top_k 20 and a presence penalty of 1.5. For precise coding it recommends temperature 0.6, the same top_p and top_k, and a presence penalty of 0[1].

Read in order, the card makes two hardware claims. The introduction says the model "runs on a single DGX". The serving section says that because it is a 122B MoE, "it serves on a single 80GB GPU", and the commands under that sentence download the BF16 repository[1]. Ollama's readme uses a third wording, "a single NVIDIA DGX card"[9]. The card also says a quantised Ornith-1.5-122B-Mobile variant is for edge devices. A public page at ornith-ai/Ornith-1.5-122B-Mobile was not available when this story was checked (the request returned HTTP 401)[1].

The BF16 files are about 250 GB, and the FP8 repository is about 133 GB. On the GGUF side, Q8_0, Q6_K and Q5_K_M are all over 80 GB. Q4_K_M, at 75.8 GB, is the GGUF file under 80 GB, before a KV cache for a long context[1][3][4]. The NVFP4 repository is about 78 GB of files and the 4-bit MLX repository about 69 GB[8][6]. The card does not say the 80GB sentence refers to those smaller builds. It also does not name which DGX "a single DGX" means. A DGX system and one 80GB GPU are different machines, and the card leaves both sentences standing.

Ornith says the model achieves state-of-the-art results among open-source models of comparable size. On its own table, Terminal-Bench 2.1 (the table spells it TerminalBench) is 81.2, behind Qwen3.8-Flash-Next at 83.4 and DeepSeek-V4-Flash-0731 at 82.7, and ahead of Beam at 80.1. SWE-bench Verified, in the column marked "non-hack", is 84.5. SWE-bench Pro "no-hack" is 76.9 and SWE-bench Multilingual "non-hack" is 76.4. CyberGym is 80.0. NL2Repo is 51.2, behind DeepSeek's 54.2 on the same table. HLE with tools is 48.0 and HLE without tools is 37.68[1].

The footnotes say Terminal-Bench 2.1 was run in Claude Code 2.1.126, averaged over five runs, and the SWE-bench scores used OpenHands at temperature 1.0, top_p 0.95 and a 256K context, with anti-hacking checks that mark a hacking solution incorrect. HLE used Claude 4.6 Opus as the judge. Those are Ornith's figures[1].

The card names no individuals. The suggested citation author is the Ornith Team[1]. The Hugging Face org profile says "We're the DeepReinforce team" and links that name to ornith.ai. The page shows 10 team members and marks the list private[2].

Two papers listed on the org name people, both under the Ornith Team. CUDA-L2 (arXiv:2512.02551, current version v4) lists Songqiao Su, Xiaoya Li, Albert Wang, Guoyin Wang, Jiwei Li and Chris Shum, with ornith.ai correspondence addresses[11][2]. GrandCode (arXiv:2604.02721, current version v3, dated 5 August 2026) lists Xiaoya Li, Guoyin Wang, Songqiao Su, Chris Shum and Jiwei Li, again with ornith.ai addresses[12]. The 122B release itself adds no further names.

Every 122B repository checked for this story sets license: mit in the model-card metadata, and the main card links LICENSE[1][3][4]. That file is not in the repository. The raw URL returns 404, and the same check on the GGUF repository also 404s. Hugging Face shows an MIT tag because the card says MIT.

For someone shipping agents, the practical offer is a commercially labelled open-weight model with a 262K context, tool-call parsing, and recipes for vLLM, SGLang, llama.cpp and Ollama. The deployment question is which file fits the machine. The 80GB sentence and the DGX sentence are both in the card, and the BF16 weights the serve commands download are about 250 GB. Match a quant to the VRAM you have, including cache, before you treat "single GPU" as a plan. And rerun the benchmarks that matter to you: the table on the card is Ornith's, including the "non-hack" SWE-bench setup.