Ornith-1.5-35B-A3B is an open-weight reasoning model for coding, repository work and agents that call tools. Its Qwen3.5-family mixture-of-experts architecture contains about 35 billion parameters but activates roughly 3 billion per token. Ornith describes training that jointly improves task generation, agent scaffolds and solutions through reinforcement learning. The published context window is 262,144 tokens. It is a candidate for developers who want to run a coding assistant on their own infrastructure: the official card documents vLLM and SGLang servers with OpenAI-compatible APIs, plus GGUF use through Ollama or llama.cpp. Start with a small repository task and check the result against your tests before expanding its role.
- Run it
- Self-hosted
- Pricing category
- free
- Publisher
- Ornith AI
Hosting and compute costs depend on how you run it.
Recorded capabilities
- Reasoning output and structured function calls with the documented runtime parsers
- Official GGUF distribution for quantized local inference
- Recent runtime versions required: Transformers 5.8.1+, vLLM 0.19.1+ or SGLang 0.5.9+
- Investigating a codebase and proposing changes for human review
- Testing local coding agents with repository and terminal tools
- Working with long documents or source files within available memory
Before choosing
- About 3B active parameters describes computation per token, not the memory needed to load the model. The publisher estimates roughly 70 GB for BF16 weights; its 256K serving example uses two 80 GB GPUs. Quantization reduces weight memory, while long contexts need additional memory.
- Coding benchmark results are publisher-reported and have not been reproduced here. Runtime, agent harness, chat template and quantization can change results.
- The model card declares MIT, but its linked licence file returned 404 on 15 September 2026; check the actual licence terms before relying on redistribution rights.
- The documented approximately 1M-token extension uses YaRN scaling; it is not the default context and may reduce quality on shorter inputs.