Overview
Ornith-1.5-35B-A3B is the mid-size model in the Ornith 1.5 family, released on August 18, 2026 under MIT. It is a sparse mixture of experts: 36 billion parameters in total, of which about 3 billion are active per token. Ornith trains it with a self-improvement loop in which the model proposes its own training tasks and learns from its own attempts.
| Specification | Ornith-1.5-35B-A3B |
|---|---|
| Total parameters | 36B |
| Active parameters | about 3B per token, 8 of 256 experts |
| Architecture | Sparse MoE, hybrid attention, built on the Qwen3.5 stack |
| Layers | 40 |
| Context window | 262,144 tokens |
| Modalities | Text and image input |
| Reasoning | Thinking trace returned separately from the answer |
| Release date | August 18, 2026 |
| License | MIT |
The sparse routing is why this one is worth the download: it costs roughly what a 3B model costs to generate a token, while scoring like a model many times larger. Memory is the price, since all 36 billion parameters still have to be resident.
Ornith-1.5-35B-A3B benchmarks
Ornith's launch numbers, from the model card, compare it with the model it replaces, its direct same-size rival Qwen3.6-35B-A3B, Muse Glimmer 30B, and the far larger Qwen3.5-397B:
| Ornith-1.5-35B-A3B | Ornith-1.0-35B-A3B | Qwen3.6-35B-A3B | Muse-Glimmer-30B | Qwen3.5-397B | |
|---|---|---|---|---|---|
Terminal-Bench 2.1 Agentic terminal-use tasks | 67.8 | 64.2 | 52.5 | 51.7 | 53.5 |
SWE-bench Verified Real-world software engineering tasks | 79.0 | 75.6 | 73.4 | 76.0 | 76.4 |
SWE-bench Pro Real-world software engineering, harder split | 59.6 | 50.4 | 49.5 | 51.2 | 51.6 |
GPQA Diamond Expert-level science questions | 89.2 | 86.2 | 86.0 | 83.5 | 88.4 |
Humanity's Last Exam Expert-level questions, no tools | 25.6 | 20.8 | 21.4 | 22.0 | 28.7 |
MCP-Atlas Tool use over MCP servers | 70.2 | 64.4 | 62.8 | 75.5 | 72.3 |
ClawEval Agentic coding in a real harness | 72.5 | 69.8 | 68.7 | - | 70.7 |
It leads five of the seven rows, and it is ahead of Qwen3.5-397B, a model ten times its size, on all three coding benchmarks. Muse Glimmer still takes MCP-Atlas, so for pure tool routing the gap runs the other way.
Ornith-1.5-35B-A3B hardware requirements
The system requirement to check is memory. We quantized the model from the original weights and published the builds as AtomicChat/Ornith-1.5-35B-A3B-GGUF.
| Memory | Build to pick | File size |
|---|---|---|
| 16 GB | AD-IQ3_XXS-IQ2_S | 13.7 GB |
| 24 GB | IQ4_XS | 18.7 GB |
| 32 GB | Q5_K_M | 24.7 GB |
| 48 GB and up | Q6_K | 28.5 GB |
Neighbouring files differ by a gigabyte or two, so when two builds both fit, take the larger one. Because only about 3B parameters are active per token, this model stays fast even when part of it sits in system RAM instead of VRAM, which is not true of a dense model the same size.
How to run Ornith-1.5-35B-A3B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Ornith-1.5-35B-A3B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
The full walkthrough, with the quant table and the llama.cpp commands, is in our guide to running Ornith 1.5 35B locally.
The smaller sibling is Ornith-1.5-9B, and both are listed on the Ornith family page. For how it compares with the rest of the local field on code, see the best local LLMs for coding.
License
Ornith-1.5-35B-A3B is released under the MIT license. It permits commercial use, modification, redistribution and private deployment as long as the copyright notice stays with the code.