Ornith-1.5-9B

// Overview

Ornith-1.5-9B is the smallest member of the Ornith-1.5 family released on 18 August 2026, a 9-billion-parameter dense model that Ornith describes as designed for efficient single-GPU deployment and edge-deployable on mobile devices through a quantised Ornith-1.5-9B-Mobile variant. In bf16 the checkpoint is about 19 GB. The weights are MIT-licensed on Hugging Face, with GGUF and MLX builds (down to 4-bit) published alongside them.

It is trained by the same end-to-end self-improvement loop as its larger siblings: rather than learning from a fixed set of human-curated tasks, the model proposes new tasks, constructs a task-specific scaffold of instructions, tools and decomposition strategies for each, and improves through reinforcement learning on the resulting solution rollouts. The base is Qwen3.5, and the model keeps that base's 262,144-token native context, which Ornith documents extending to roughly one million tokens with YaRN RoPE scaling at a factor of 4.0.

Ornith publishes a full serving quickstart: `vllm serve ornith-ai/Ornith-1.5-9B` with the qwen3 reasoning parser and qwen3_xml tool-call parser, an equivalent SGLang launch, `ollama run ornith-1.5:9b`, and `llama-server -hf hf.co/ornith-ai/Ornith-1.5-9B-GGUF`. Recommended sampling differs by task — temperature 1.0 with presence_penalty 1.5 for general use, temperature 0.6 with no presence penalty for precise coding.

On Ornith's published evaluations the 9B model reaches 70.6 on SWE-bench Verified and 86.4 on GPQA Diamond, well ahead of the similarly-sized Qwen3.5-9B (53.2 and 81.7) and of the much larger dense Gemma-4-31B on the coding rows (52.0 on SWE-bench Verified). It generally trails the ~35B-class Qwen3.6-35B-A3B on breadth rows such as MCP-Atlas and BrowseComp while beating it on SWE Atlas QnA and NL2Repo.

// Benchmarks

This model's scores

1. GPQA Diamond86.4%
2. SWE-bench Verified70.6%
3. ClawEval66.5%
4. WideSearch59.5%
5. BrowseComp56.4%
6. SWE-bench Multilingual54.4%
7. MCP-Atlas54.2%
8. SWE-bench Pro47.5%
9. Terminal-Bench 2.1 (Claude Code)47%
10. Toolathlon-Verified41.2%
11. NL2Repo32.4%
12. HLE (with tools)30.5%
13. SWE Atlas – QnA20.6%
14. HLE (no tools)20.2%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• MIT-licensed 9B weights that fit a single GPU (~19 GB in bf16), with GGUF and 4-bit MLX builds for laptops
• SWE-bench Verified 70.6 and GPQA Diamond 86.4 on Ornith's evaluations — far ahead of the same-size Qwen3.5-9B
• Keeps the family's 262K native context, documented to extend to roughly 1M tokens with YaRN
• One-line launches for vLLM, SGLang, Ollama and llama.cpp, with agent tool-call and reasoning parsers configured
• Published per-task sampling recommendations (general vs precise coding) rather than a single default

// Best for

• Local and on-device coding assistants where the whole model must fit one GPU or a laptop
• Edge and mobile deployment through the quantised Mobile and MLX builds
• Cheap high-volume agent steps in a pipeline that escalates hard cases to a larger model
• Fine-tuning a small, permissively-licensed agentic base for a narrow tool-use domain

// Ornith (open-weight) — every version

The full lineage of the Ornith (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.