Ornith 1.5
Agentic-coding models post-trained from Qwen3.5 in 9B, 35B-A3B, 122B and 397B sizes, MIT licensed. All benchmark claims are self-reported.
In short. Ornith 1.5 is a family of open AI models tuned to work as coding agents: they read a codebase, run commands, fix bugs and call tools. They are post-trained versions of Alibaba's Qwen3.5 in four sizes, from a 9B model for a single GPU to a 397B model for a multi-GPU server. All use the MIT license.
Summary
Ornith is made by DeepReinforce, a California AI research startup that was earlier known for reinforcement learning on GPU code. Ornith 1.0 came out in June 2026, built on Qwen3.5 and Gemma 4 with continued pretraining, mid-training and reinforcement-learning post-training. Ornith 1.5 followed on 19 August 2026 with a 9B dense model and 35B-A3B and 397B mixture-of-experts models, and a 122B mixture-of-experts model was added in October. All four checkpoints have a 256K-token context and think in a reasoning block before they answer. The config files of the 35B-A3B, 122B and 397B models match Qwen3.5-35B-A3B, Qwen3.5-122B-A10B and Qwen3.5-397B-A17B, which activate about 3B, 10B and 17B parameters per token. Ornith itself states only the 3B figure.
DeepReinforce's training method lets the model propose its own training tasks, write a scaffold of instructions, tools and orchestration for each one, and then solve it, with reinforcement learning rewarding all three steps. For the 397B model the blog reports 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, against 85.0 and 59.0 for Claude Opus 4.8, and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731. Kimi K3, at 88.3 and 67.5, scores higher in the same table. For the 35B-A3B model it reports 67.8 on Terminal-Bench 2.1 and 79.0 on SWE-bench Verified, against 52.5 and 73.4 for Qwen3.6-35B-A3B. The 9B model scores 46.2 and 70.6.
All results are Ornith's own, averaged over five runs, and the cards say the Qwen chat template was adjusted for the Terminal-Bench runs. The 122B card shows Qwen3.8-Flash-Next ahead on Terminal-Bench 2.1, 83.4 against 81.2. Serving needs Transformers 5.8.1 or newer, vLLM 0.19.1 or newer, or SGLang 0.5.9 or newer, with the reasoning and tool-call parsers the cards list. Ornith also publishes FP8, NVFP4, GGUF, MLX and DFlash builds. On 11 November 2026 Hugging Face showed about 5.7 million downloads over 30 days for the 9B GGUF build and 4.2 million for the 35B-A3B GGUF build.
Good for
- Coding agents that run commands and edit repositories, locally with the 9B or 35B-A3B model
- Teams that want an MIT-licensed model on the Qwen3.5 architecture, with FP8, NVFP4, GGUF and MLX builds
- Tool calling through an OpenAI-compatible server with the Qwen-style tool-call parsers
- Long tasks that need up to 256K tokens of context
Watch out for
- Every benchmark number comes from Ornith, including its runs of competing models, and some of its own tables put rivals ahead
- It is a post-trained derivative of Qwen3.5, and the published scores focus on coding and agent tasks
- It thinks before answering by default, which costs tokens, and the cards recommend temperature 0.6 for coding and 1.0 for general tasks
- The 397B model needs about 807 GB in BF16 or 418 GB in FP8, so it needs a multi-GPU server
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Ornith-1.5-9B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Ornith-1.5-9B BF16 | 9.65B | 22 GB | 11 GB | 6.66 GB | Laptop or phone |
| Ornith-1.5-9B (NVFP4) NVFP4 | 9.65B | 22 GB | 11 GB | 9.7 GB | Laptop or phone |
| Ornith-1.5-35B-A3B BF16 | 36B 3B active | 83 GB | 41 GB | 25 GB | Gaming GPU |
| Ornith-1.5-35B-A3B (FP8) FP8 | 36B 3B active | 83 GB | 41 GB | 25 GB | Gaming GPU |
| Ornith-1.5-35B-A3B (NVFP4) NVFP4 | 36B 3B active | 83 GB | 41 GB | 26 GB | Gaming GPU |
| Ornith-1.5-122B BF16 | 125B 10B active | 288 GB | 144 GB | 86 GB | Workstation |
| Ornith-1.5-122B (NVFP4) NVFP4 | 125B 10B active | 288 GB | 144 GB | 85 GB | Workstation |
| Ornith-1.5-397B BF16 | 403B 17B active | 928 GB | 464 GB | 278 GB | Big-memory desktop |
| Ornith-1.5-397B (FP8) FP8 | 403B 17B active | 928 GB | 464 GB | 278 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Ornith 1.5 vs Qwen3.6: which agentic coding model should you use?
- Best open-source models for coding and AI agents
- Hydrogen-1.0: a 753B open science model built on GLM-5.3
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is Ornith 1.5?
- Ornith 1.5 is a family of open AI models tuned to work as coding agents: they read a codebase, run commands, fix bugs and call tools. They are post-trained versions of Alibaba's Qwen3.5 in four sizes, from a 9B model for a single GPU to a 397B model for a multi-GPU server. All use the MIT license.
- Can I use Ornith 1.5 commercially?
- Yes. Ornith 1.5 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Ornith 1.5 open source?
- Ornith 1.5 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Ornith 1.5 on my own computer?
- Yes. The smallest variant, Ornith-1.5-9B, needs about 6.66 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Ornith 1.5 released?
- Ornith 1.5 was released by DeepReinforce on August 19, 2026.
- How long a context window does Ornith 1.5 have?
- Ornith 1.5 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Ornith-1.5: From Self-Scaffolding to Self-Improvement (Ornith blog, Aug 2026) ornith.ai
- Ornith repository and checkpoint table (GitHub) github.com
- Ornith-1.5-35B-A3B model card (Hugging Face) huggingface.co
- Ornith-1.5-122B model card (Hugging Face) huggingface.co
- Release report with the 19 August 2026 date (GIGAZINE) gigazine.net
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Muse Glimmer Meta | 29.6B | 128K | Apache 2.0 (Permissive) | 20 GB Gaming GPU |