Ornith 1.5 35B A3B

KGP Talkie runs a hands-on benchmark of ornith-ai's Ornith-1.5-9B-GGUF and Ornith-1.5-35B-A3B-GGUF on a single RTX 5090 with 32 GB VRAM, using llama-server build 10448 on Windows 11, Q4_K_M quantization for all models, f16 KV cache with flash attention, and 512-token decodes via the raw /completion endpoint with ignore The article walks through reading Ornith's hybrid block layout, why decode speed tracks active parameters (and why that isn't unique to Ornith), KV-cache cost per token across the three architectures, and why 'max context that loads' on Windows is a misleading figure. It also notes two practical pitfalls: the --jinja f

A note.com deep dive (auto-translated from Japanese by AI-Driven Lab) reports that DeepReinforce released the open-weight LLM 'Ornith-1.5' on August 19, 2026, and that the family ships in three sizes — 397B, 35B, and 9B. The article describes Ornith-1.5's headline mechanism as a 'self-improvement loop' in which the mod The piece assumes intermediate familiarity with MoE and RL concepts and is targeted at engineers and product managers evaluating LLM selection. Because it is an AI-translated, third-party explainer rather than a first-party release note, the 'self-improvement loop' claim and the DeepReinforce attribution should be trea

On August 19, 2026, DeepReinforce released the Ornith-1.5 family of open-weight models in 9B, 35B-A3B, and 397B sizes across multiple quantizations and formats, as reported in a community thread on the NVIDIA DGX Spark / GB10 developer forum. The Ornith-1.5-35B-A3B variant was highlighted as well-suited for single-node The forum report cites the Ornith model card's sampling recommendations (temperature 0.6 for general tasks, 1.0 for benchmarking) and notes the model's text-only design, distinguishing it from multimodal peers like Qwen3.6-35B. The model card states the context window can be pushed to 1M tokens via RoPE, though the tes

Atomic Chat published a detailed how-to guide on August 19, 2026 for running Ornith 1.5 35B locally, describing it as the mid-size member of DeepReinforce's new Ornith 1.5 family and the strongest variant that fits a desktop. The page lists concrete specs taken from the model card: 34.7B total parameters plus a 1.9B Mu The guide further documents Atomic Dynamic GGUF builds and hardware requirements, noting the model fits a 24 GB GPU or a 32 GB Mac at 4-bit quantization, and walks through setup with both Atomic Chat and llama.cpp. It attributes creation to DeepReinforce rather than RunInfra and makes no gateway or hosting claims, keep

A BenchLM benchmark aggregation page for Ornith-1.5-35B-A3B (data as of September 23, 2026) reports a capability score of 30.6/100 against a field median of 50.4, placing it 156th of 196 ranked models. The page tracks 18 published benchmark rows and documents a 262K token context window, with pricing listed as self-hos Category-level rankings show the model at rank 84/105 (20th percentile) for Agentic across 7 verified benchmarks, rank 104/135 (23rd percentile) for Coding across 7 verified benchmarks, and rank 121/160 (25th percentile) for Knowledge across 4 verified benchmarks. Reasoning, Multimodal, Multilingual, Instruction-follow