DeepReinforce
Ornith 1.0 is DeepReinforce's self-improving open-source model family for agentic coding. The Hugging Face organization and model cards list 9B-Dense, 31B-Dense, 35B-MoE and 397B-MoE positioning, with public 9B, 35B and 397B checkpoints plus GGUF and FP8 variants. The cards say the models are post-trained on top of Gemma 4 and Qwen 3.5, use reinforcement learning to optimize both solution rollouts and the scaffold that drives those rollouts, and target coding-agent benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw / ClawEval. The model cards document reasoning-model behavior, OpenAI-compatible serving through vLLM or SGLang with Qwen-style reasoning and tool-call parsers, Transformers loading, 262K context serving examples for 9B/35B/397B, MIT licensing, and globally accessible weights.
Editorial verdict
Teams evaluating open agentic-coding models for self-hosted coding agents, terminal automation, SWE-Bench workflows and long-context tool use.
Avoid it if you need a polished hosted coding IDE, managed team billing or a small model that runs comfortably without GPU planning.
Ornith 1.0 belongs in AI Coding because its public model cards and collection position it around agentic coding benchmarks and coding-agent deployment.
MIT open weights on Hugging Face; self-hosted inference costs depend on model size and quantization
Hugging Face model download, Transformers, vLLM, SGLang, GGUF local inference
Commercial use should follow the current product, API, model license and billing terms.
Review prompt, file, media upload, retention and training-use terms before sensitive workloads.
Serve Ornith through vLLM or SGLang with reasoning and tool-call parsers for coding-agent experiments.
Reproduce Terminal-Bench, SWE-Bench, NL2Repo, SWE Atlas or ClawEval results before selecting a model size.
Evaluate 9B or GGUF variants for lighter experiments, and FP8 variants when compressed large-model serving matters.
Model names, quotas, release status, regional access and commercial terms can change quickly; recheck official sources before procurement or production use.
qwen-agentworld
qwen-code
kimi-k2-7-code
minimax-m3
zhipu-glm
official · en · verified 2026-08-04
Confirms the DeepReinforce org, website, Ornith-1.0 collection, seven public model repositories, datasets and organization card positioning.
official · en · verified 2026-06-26
Confirms the 397B MoE model card, MIT license, agentic-coding positioning, benchmark disclosures and vLLM/SGLang serving recipes.
official · en · verified 2026-06-26
Confirms the 35B variant, benchmark table, reasoning model behavior and deployment requirements.
official · en · verified 2026-06-26
Confirms the 9B dense variant and single-GPU-oriented local serving guidance.
docs · en · verified 2026-06-26
Linked from the official model cards as the Ornith blog path.
Last checked: 2026-08-04