DeepReinforce (MIT-licensed agentic coder)

Ornith 1.0

Local agentic coder that learned to write its own task harness — the small one for daily work.

Context262,144 tokens (Qwen35 family, 9B dense)

PricingFree · runs locally

Tasks tested15

Avg scorecurrently unranked

Medals🥇0 🥈0 🥉0

Release2026-06-25

Official sitehuggingface.co/deepreinforce-ai ↗

Official vendor source

Ornith 1.0 is built by DeepReinforce (MIT-licensed agentic coder) — see the vendor's own product page, pricing, and docs at huggingface.co/deepreinforce-ai.

Visit huggingface.co/deepreinforce-ai →

What is Ornith 1.0?

Ornith 1.0 is the DeepReinforce (MIT-licensed agentic coder) frontier model with a 262,144 tokens (Qwen35 family, 9B dense) context window, released 2026-06-25. Tagline: Local agentic coder that learned to write its own task harness — the small one for daily work.. Official source: huggingface.co/deepreinforce-ai.

Pricing detail. Ornith-1.0-9B is the local end of DeepReinforce's MIT-licensed Ornith family. Q8_0 GGUF is 9.5GB. Runs 100% offline on a consumer Mac via Ollama. The flagship 397B variant (cloud-only, no API host yet) self-reports 82.4% on SWE-Bench Verified — promising but not independently verified. The 9B is the daily-driver local end.

How I use it inside the Agent OS. GLOBAL local model for all of Agent OS — Local chat, Local Hermes Engine, Agent Kanban (Planner→Builder→Reviewer), loop judge. One setting (LOCAL_MODEL=ornith-9b) drives every local surface.

What I built with Ornith 1.0

Every model on Goldie Bench gets the same fixed prompt set — one shot, single HTML file out — and I score the result 0–10 inside the Agent Operating System. Here's what Ornith 1.0 shipped on the bench: 15 one-shot demos across 262,144 tokens (Qwen35 family, 9B dense) of context. Of those, 0 are scored against the field with my honest 0–10 from the source guides at agentos.guide.

Strengths

Real agentic — tools + thinking capabilities (not just chat) inside a 9B local model
~36 tokens/sec warm on M4 Max — same ballpark as Gemma's local coder
Self-fixed its own frame-rate bug on a Snake build when handed back the broken output

Trade-offs

9B parameter ceiling — not in the same league as 100B+ frontier models for complex 3D / WebGL
Output wrapped in multiple markdown fences + prose preamble — needs structured extraction (the bench harness does this)

Best for

Free, private, offline coding on a consumer Mac with real tool-calling
Daily local builds where 9B is enough and you don't want the per-token bill
Agent OS local-fallback when frontier API access is offline or rate-limited

Every demo by Ornith 1.0

15 live demos, sorted by category. Click any tile to play the actual one-shot result. Verdicts and 0–10 scores are pulled from the source guides where I posted them publicly.