AI Potluck
Model components / Fine-tuned / chat models

Ornith 1.0 9B

DeepReinforce

DeepReinforce's open-weights agentic coding model, a 9B dense model built on the Qwen 3.5 architecture, with explicit <think> reasoning and tool-calling for terminal-based software-engineering tasks. Optimized for single-GPU deployment (~19GB in bf16).

MIT-licensed open weights (HF tag license:mit), Qwen 3.5 architecture (qwen3_5), released 2026 by DeepReinforce. Reported coding benchmarks in the card's model-index are SWE-bench Verified 69.4, SWE-bench Pro 42.9, SWE-bench Multilingual 52.0 and Claw-Eval 63.1; the Terminal-Bench figure the record used to quote is not in the model-index. The repos are now served under the org ornith-ai; the declared deepreinforce-ai URLs still resolve. Verified 2026-08-13 via the HF model cards.

Openness

3 high confidence
3.0
weights
open(MIT)
base
Qwen3.5
data
closed
recipe
closed
license
MIT(OSI)

MIT-licensed open weights on a Qwen 3.5 base, but no open training data or recipe, so open weights rather than open source. The repository is now served under the org ornith-ai rather than deepreinforce-ai; the declared URL still resolves.

Adoption

4 high confidence
4.0

Banded on the sum of Hugging Face downloads across the two declared repositories: 2,383,936 for Ornith-1.0-9B and 4,594,817 for Ornith-1.0-9B-GGUF, summing to 6,978,753 in the trailing 30 days, which lands in 1M-10M. Both figures come from the Hugging Face API rather than the rendered card. The two repositories carry 528 and 638 likes.

Capability

4 medium confidence
4.0

Strong agentic-coding scores for a 9B model and competitive within the category, but below the frontier coders; the benchmarks are vendor-reported on the model card. That card's model index also carries SWE-bench Multilingual 52.0 and Claw-Eval 63.1. The Terminal-Bench figures quoted here are not in the model index and cannot be re-derived from it.

Verified 2026-08-13