DeepReinforce's first release is a family of agentic coding models, AI systems that write and edit code through long multi step loops, built on Google's Gemma 4 and Alibaba's Qwen 3.5 and claiming top coding benchmark scores among comparable open models — trained to build its…
The Ornith-1.0 model family, released this past week from a previously low-visibility lab called DeepReinforce, is one of the more interesting entries in this year's open-weights coding-model cycle, both because of how it was trained and because of how little the people behind it have put on the public record.
The headline-level story is straightforward: a downloadable, MIT-licensed set of model weights, sized to run on a single machine. The 9B and 31B dense variants sit alongside 35B and 397B mixture-of-experts variants, with the 35B Q4_K_M GGUF weighing in around 20GB, roughly the size a capable laptop-class machine with offloaded layers can chew through. DeepReinforce's release post and the official site frame the family as the lab's first model release, distributed openly under terms compatible with the Apache 2.0 Gemma 4 and Qwen 3.5 base models it is built on top of.
The mechanism DeepReinforce is highlighting is what makes the release worth a closer read. Most agentic-coding models are evaluated and trained against agent harnesses built by humans: small programs that loop the model through tool calls, scratchpads, retries, and planning steps the researchers have hand-engineered. DeepReinforce calls its approach self-scaffolding, which in plain language means the model is trained to discover the scaffolding for that loop on its own rather than inheriting one designed by the lab. The pitch is that a model optimized against its own evolving tool-use policy should generalize better to fresh, real coding workflows than a model measured only against a fixed harness.
Whether the claim holds is the load-bearing question, and the source basis is mostly DeepReinforce's own at this point. The lab says Ornith-1.0 leads comparable-size open models on standard coding benchmarks, but most of those numbers come from the release material itself until independent leaderboard entries and third-party reruns land. What already does exist outside the lab is practitioner signal. Simon Willison ran the 35B Q4_K_M GGUF locally and reported capable multi-step tool-call behavior against a real Datasette codebase, with the same setup producing roughly 103 tokens per second on a local image-generation workload. A Classmethod developer-community post running on a DGX Spark cluster benchmarks the 35B variant including Japanese-language coding tasks, and a thread on the NVIDIA developer forums is already collecting inference-tuning notes for the 397B model. Those are community data points, not leaderboard rankings, but they are the kind of evidence that tells a reader whether a model actually behaves like an agent on real code rather than only on test items.
Practically, the path to running it tonight is short. The 9B checkpoint is on Hugging Face, as is the 35B GGUF and the full 397B MoE model, with the GitHub repository hosting the open-source code. Willison's local setup uses LM Studio with the Pi agent harness, which is a reasonable default. Readers with a DGX Spark or comparable high-memory workstation can pull the 397B directly.
That is also where the honest gap sits. DeepReinforce has a June 2025 paper on contrastive reinforcement learning for CUDA optimization, called CUDA-L1, the lab's only clearly traceable prior artifact and a reasonable pedigree for a self-scaffolding RL story. Beyond that, the lab's public footprint is thin: no well-known funding announcement, no large customer roster, no executive track record a curious reader can cross-reference. The MarkTechPost coverage of the release treats it as news precisely because so little is on the record yet. For most open-weights drops that opacity would be a passing footnote. For a model claiming state-of-the-art coding performance via a novel training method, it is the part a reader should hold at arm's length until independent reruns land.
What to watch next is narrow. Third-party coding-leaderboard entries, especially on long-horizon agentic coding benchmarks like SWE-Bench Verified where the scaffolding question should show up in the results, will settle whether the self-scaffolding claim survives outside DeepReinforce's evaluation harness. A peer-reviewed version of the CUDA-L1 lineage, or a confirmed employer or investor note for the lab, would also move the credibility dial. Until then, the right frame for a reader is pragmatic. Ornith-1.0 is a real, downloadable, MIT-licensed open coding model from a known-quantity base-model chain, and the open question is whether DeepReinforce ships a second checkpoint.