Long agentic coding workflows need models that (1) call tools reliably, (2) expose internal reasoning traces safely, and (3) handle very long contexts. Ornith-1.5-35B-A3B-GGUF packages a 35B mixture-of-experts checkpoint into a GGUF build designed for local serving; its training emphasized a self-improvement loop that jointly optimizes task generation, scaffold construction, and solution rollouts, producing strong agentic and coding performance in the authors' evaluations.
Key Capabilities
- Long-context support: native context up to 262,144 tokens (authors validate YaRN rope-scaling to extend toward ~1M tokens). So what: you can run retrieval-augmented or multi-file code reasoning without chopping inputs into many short windows.
- Mixture-of-experts economics: ~35B total params but only ~3B activated per token (A3B configuration). So what: throughput/latency behave like a ~3B model while memory and storage must accommodate the full expert set.
- Reasoning + tool-calling: model emits a separate
<think>trace and well-formed function/tool-call blocks parsed into OpenAI-style tool_calls. So what: integrates cleanly with agent frameworks and lets servers surface chain-of-thought separately from final answers. - Agentic coding focus and benchmark results: authors report sizable gains on Terminal-Bench, SWE-bench, NL2Repo and agentic benchmarks versus several dense peers. So what: a strong candidate when you need automated code editing, repo understanding, or multi-step tool orchestration locally.
Who it's for and tradeoffs
Great fit if you run local or private agentic coding workloads that need long context windows, explicit tool orchestration, and isolated local inference (GGUF/llama.cpp/vLLM paths). Look elsewhere if you lack large GPU memory: MoE models require all experts resident (authors recommend multi-80GB GPUs) and impose higher memory/storage costs than similarly-sized dense models. Also note the published benchmark numbers are the authors' reports; independent reproduction may vary and quantized GGUF builds trade a small amount of fidelity for much smaller disk footprints.
Where it fits
Compared to dense 31–35B models, Ornith-1.5-35B-A3B targets agentic coding and tool use with an MoE cost/performance tradeoff — lower activation footprint per token but higher total memory. If you need compact, single-GPU deployment with minimal memory, a smaller dense model or heavily quantized build may be easier; if you need scalable agentic performance and long context, Ornith's MoE design is compelling.
