Ornith 1.5

Using the V100 (16GB VRAM) as a reference, the Ornith 9B model at Q6_K requires approximately 7–8GB of VRAM, while the 35B MoE requires approximately 20–22GB and needs dual V100s.

In a test of 30 Three.js One-Shot programming problems, Qwen 3.8 passed 24 and Ornith 1.5 passed 15, both running under Q6 quantization.

Qwen3.8-27B加载后占用16.1GB内存，Ornith 1.5占用19.8GB内存。

在800 token输入提示下，Qwen3.8-27B的首token延迟为3.46秒，Ornith 1.5为0.49秒，后者的首token延迟约为前者的七分之一。

Qwen3.8-27B的输出速度为55.1 tokens/秒，Ornith 1.5为90.4 tokens/秒。

Qwen3.8-27B是稠密模型，全部270亿参数参与每个token的计算，依靠MTP头预测后续token并在单次前向中验证；Ornith 1.5采用MoE架构，每token仅激活约30亿参数，配合DFlash由小型草稿模型提议token块再由主模型校验。

在关闭推理模式的情况下，Qwen3.8-27B在20个植入bug中检出9个，Ornith 1.5检出8个，差异属于统计噪声。

开启推理模式后，Qwen3.8-27B的bug检出率降至6/20，40个回答中有12个因耗尽12000 token输出上限导致JSON无法解析；Ornith 1.5检出率维持8/20，13/40回答不可解析，但耗时约为Qwen的一半。

Ornith 1.5 adopts a three-stage reinforcement learning architecture, in which the first-stage task generator autonomously analyzes codebases to generate training tasks and receives rewards based on correctness, difficulty, and novelty.

In a 3D game building test on an Apple M3 Ultra with 512GB unified memory, Qwen 3.8 completed it in one shot, while Ornith 1.5 required 6 separate iterations to barely approach the same level.

In an ultra-long context test, Ornith 1.5 ran at around 75 tokens/second early on and dropped to approximately 45 tokens/second after exceeding 1.6 million tokens, while Qwen 3.8 dropped to approximately 5 tokens/second under the same load.

Both models are supported by llama.cpp, Ollama, and LM Studio, and both support MTP multi-token prediction speculative decoding acceleration.

Ornith 1.5 35B is a Mixture-of-Experts (MoE) model with 35 billion total parameters, activating approximately 3 billion per token, containing 256 independent experts internally, with a routing mechanism that selects 8 of them to process each token.

Ornith 1.5 uses the MIT license and is one member of a three-model family that includes a 9B dense version, a 35B MoE version, and a 397B flagship version.

In a task to generate a 3D game from scratch, Qwen 3.8 completed it with a single prompt and zero guidance, while Ornith 1.5 required six iterations.

Ornith 1.5 requires approximately 23 GB of VRAM for the recommended Q4_K_M quantization with the full package, and it is advised to prepare 27 GB or more; Qwen 3.8 at the same Q4_K_M tier requires only 16 to 19 GB.

The Ornith 1.5 family includes a 9B compact version and a 397B flagship version, while Qwen has a Max version with 2.4 trillion parameters available via API.

Ornith 1.5 uses the MIT license and can be freely modified and used commercially.

Ornith 1.5 is a Mixture of Experts (MoE) model with 256 experts internally, activating only 8 experts per token, with approximately 3B parameters actually activated.

The routing in a MoE model dynamically selects experts on a per-token basis, so all 35B parameters must reside in VRAM simultaneously—computation is cheap but VRAM usage is not reduced.