O

Ornith-1.0

DeepReinforce · Released Jun 2026

Conditional

A strong MIT-licensed open-source agentic coding model family that achieves frontier-competitive benchmarks — beating Claude Opus 4.7 on both Terminal-Bench 2.1 and SWE-Bench Verified — but requires self-hosting with significant GPU resources and comes from a new lab with limited track record.

Is it right for you?

Good for

  • Agentic coding with multi-step tool calling — outperforms Claude Opus 4.7 on Terminal-Bench 2.1 (77.5 vs 70.3) and SWE-Bench Verified (82.4 vs 80.8)
  • Self-hosted coding agents with full data privacy — MIT license, no API dependency, runs on your own infrastructure
  • Edge deployment with the 9B variant — matches or exceeds Gemma 4-31B on SWE-Bench Verified (69.4 vs 52) while fitting on a single consumer GPU
  • Self-scaffolding RL training produces per-task strategies that outperform fixed human-designed harnesses on agentic coding benchmarks

Not good for

  • Teams without GPU infrastructure — no managed API from DeepReinforce; you must provision and maintain your own inference servers
  • General-purpose chat or non-coding tasks — heavily specialized for agentic coding harnesses; not benchmarked or tuned for conversation, writing, or multimodal use
  • Production deployment with reliability SLAs — from a new lab with no managed service, limited community adoption data, and unverified benchmark claims

How it performs by task

Agentic coding (multi-step tool use)

Excellent

SOTA among open-source models; 397B beats Claude Opus 4.7 on both TB-2.1 and SWE-Bench Verified. Simon Willison independently verified proficient multi-tool-call behavior.

Code generation

Very Good

Strong on SWE-Bench Verified (82.4) and NL2Repo (48.2). Competitive with frontier closed models but trails Claude Opus 4.8 and GLM-5.2 on raw generation benchmarks.

Code debugging and bug fixing

Very Good

Solid SWE-Bench Pro score (62.2) and strong SWE Atlas results. Handles multi-file bug localization and test-driven patches well.

Multi-file refactoring

Very Good

Designed for repository-scale tasks. 256K context window enables working across large codebases. Strong on SWE-Bench Multilingual (78.9).

Local/edge inference

Excellent

The 9B variant (69.4 on SWE-Bench Verified) punches far above its weight class, matching 31B-class models. GGUF quants available for llama.cpp and Ollama.

Pricing

Input

Free (MIT license)

Output

Free (MIT license)

Context

256K tokens

View full pricing

Benchmarks

BenchmarkScoreSource
SWE-Bench Verified82.4 Source
Terminal-Bench 2.1 (Terminus-2)77.5 Source
SWE-Bench Pro62.2 Source
SWE-Bench Multilingual78.9 Source

No verdict changes yet

The clock starts day one — changes land here as our verdict evolves.

Verification log

  • Pricing— No changes

    Automated agent

  • Pricing— No changes

    Automated agent

  • Pricing— No changes

    Automated agent

  • Profile— No changes

    Imported at launch

  • Pricing— No changes

    Imported at launch

How we evaluate