Ornith-1.5-9B centers on delivering a small, deployable reasoning model that preserves agentic and coding capabilities without requiring very large clusters. The core insight: by coupling self-improvement training (task generation, scaffold construction, and rollout optimization) with a compact 9B architecture, the model achieves competitive agentic and coding performance while remaining runnable on a single high-memory GPU and in quantized form for edge devices.
What Sets It Apart
- Self-improvement training loop: the family is trained to generate tasks, discover scaffolds, and optimize rollouts, which shifts effort from fixed human-curated tasks to automated curriculum generation — this explains why a 9B configuration shows unexpectedly strong agentic behavior compared with older 9B baselines.
- Engineering for deployment: the dense 9B checkpoint is ~19 GB in bf16 and is distributed in GGUF/quantized variants (4-bit/8-bit) suitable for single 80GB GPU serving or mobile edge inference, reducing the hardware barrier for agentic workflows.
- Long-context and tool-friendly: native support for very large context windows (262,144 tokens by default, YaRN rope-scaling validated to extend up to ~1M tokens) and explicit tooling/tool-call outputs and a separate reasoning_content trace to preserve chain-of-thought for downstream tool parsers.
- Benchmarks that matter: strong scores on terminal-oriented coding and agentic evaluations (e.g., Terminal-Bench ~46.2, SWE-bench Verified ~70.6, MCP-Atlas ~54.2, ClawEval ~66.5) — so what: these numbers indicate the model is tuned for real-world, multi-step coding and agent workflows rather than only single-turn chat metrics.
Who it's for and tradeoffs
Great fit if you need an on-prem / local agent or coding assistant that: runs on a single high-memory GPU, needs long-context reasoning (large codebases, long dialogues, or tool-using agents), and benefits from an explicit chain-of-thought output for debugging or tool orchestration. Look elsewhere if you require absolute top-tier single-turn conversational quality or SOTA few-shot reasoning on benchmarks dominated by much larger models — larger 35B+ models still lead on some metrics. Also note it expects modern runtimes and parsers (vLLM/SGLang, reasoning/tool parsers) to expose its reasoning_content and tool_calls cleanly.
Where it fits
Use it as the inference backbone for terminal coding CLIs, local agent stacks, research into self-improving training loops, or edge deployments where a quantized GGUF build is important. It sits between tiny local models (which lack agentic depth) and multi‑GPU giant models (which are expensive to serve).
Practical considerations
The model is packaged in HF with GGUF and bf16 checkpoints and is distributed under an MIT license. Serving and integration scenarios benefit from OpenAI‑compatible endpoints and explicit tool-call support, but operational users should plan for a recent runtime stack and adequate GPU memory for best results.
