Monday, August 24, 2026
MIT-licensed weights; Hugging Face ornith-ai/Ornith-1.5-397B created 2026-08-18. Family is 397B MoE, 35B MoE (~3B active), and 9B dense. On the authors’ table, Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 vs Claude Opus 4.8 at 85.0, and 56.0 on DeepSWE vs 59.0. Context 262,144 tokens natively; YaRN factor 4.0 documented for ~1M. The 9B is described as deployable on iPhone/Android when quantized. vendor
Hybrid Mamba-2 + MoE + Attention, 30B total / 3B active, up to 1M context, OpenMDW-1.1, commercial use. NVFP4 card lists single-GPU on 1× DGX Spark (GB10) or 1× H100, plus MTP, DSpark, and DFlash. NVIDIA reports up to 4× output speed vs similar-sized models and, on PinchBench, 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy. vendor
2.4B VLM (400M native-resolution SigLIP-2-derived encoder + 2B North Micro LLM). Training preserved A4-page aspect ratio up to 1654×2339 (200 dpi). Card rows include DocVQA VAL 0.921, ChartQA Test 0.808, OCRBench 0.792. Public vLLM support listed as coming soon; MLX-VLM quants and NVIDIA AutoModel / Axolotl recipes are linked. vendor
Saturday already covered Pro GA on app/web/API. The matching Hugging Face checkpoint is MIT-licensed, supersedes the preview, and attaches a DSpark speculative-decoding module. Official GA numbers include HLE 42.7 / 60.0 (no tools / with tools), Terminal Bench 2.1 87.9, DeepSWE 62.7. vLLM and SGLang recipes enable DSpark from the same checkpoint. vendor
