On August 19, 2026, a small lab called DeepReinforce — operating under the Ornith AI banner — published the open-weights release of the summer. The Ornith-1.5 family ships under the MIT license in three sizes: a 9B dense model, a 35B mixture-of-experts with just 3B active parameters per token, and a 397B MoE flagship. The headline is not the parameter counts. It is how the models were trained: in an end-to-end self-improvement loop where the model proposes its own training tasks, builds the scaffolds to solve them, and generates the solution rollouts that reinforcement learning then optimizes.
In other words, the model was its own teacher. The weights are now yours to download.
The family
Ornith-1.5 comes in three sizes:
- Ornith-1.5-9B — dense 9B model, roughly 19GB in BF16, the one that runs on a single GPU or a beefy laptop. A quantized 9B-Mobile variant targets phones and edge devices.
- Ornith-1.5-35B-A3B — the mid-size MoE: 34.7B total parameters, about 3B active per token (8 of 256 routed experts plus one shared expert), 40 layers, hybrid attention. Roughly 70GB in BF16.
- Ornith-1.5-397B — the flagship MoE, needing around 800GB in BF16 and a multi-GPU node. Open weights, not open to your laptop.
All three share a 262,144-token native context window, extensible toward a million tokens with YaRN scaling, and a reasoning mode that thinks in a visible block by default but can be switched off. Day-one quantization builds shipped in FP8, GGUF, MLX, and NVFP4, with support across Ollama, llama.cpp, LM Studio, vLLM, and SGLang.
The self-improvement loop
Ornith-1.0, released in June 2026, generated its own task scaffolds and solutions for a fixed set of coding problems. Version 1.5 closes the loop: it jointly optimizes task generation, scaffold construction, and solution rollouts. The model proposes new training tasks, writes the harnesses that grade them, produces solution rollouts, and the whole cycle is optimized with GRPO — a reinforcement learning algorithm. Task rewards combine validity, frontier difficulty (targeting roughly 20% success rates), and novelty; the harness itself gets rewarded for fidelity and resistance to hacking.
Two clarifications matter. First, ‘self-improving’ describes the training process, not something the model does after you download it — your copy does not get smarter on your machine. Second, the lineage is continued pretraining and post-training on top of Qwen3.5 and Gemma4 bases, not a from-scratch build. Open weights here means open artifacts of a derivative training run, which is worth knowing before you reason about what the model ‘knows.’ Derivative or not, the training recipe itself is the contribution: if self-improving loops keep working, the next Ornith may not need a frontier base at all.
The benchmark claims — read carefully
DeepReinforce describes the flagship as performing comparably to Claude Opus 4.8 across reasoning, agentic workflows, and instruction following. The published numbers for the 35B-A3B variant include 86 on SWE-Bench Verified, 65.1 on SWE-Bench Pro, 79.6 on SWE-Bench Multilingual, 86.1 on Terminal-Bench 2.1, and 44.6 on Humanity’s Last Exam.
Read the fine print before repeating the headline. Independent analysis of Ornith’s own comparison table found the flagship winning three of five head-to-head benchmarks against Opus 4.8 and losing two — ‘beats Claude’ is an overstatement; ‘comparable’ is the defensible word. Against Z.ai’s GLM-5.2, the wins are cleaner. And every figure is vendor-reported: independent verification had not caught up at publication time, with at least one community benchmark of the 35B showing meaningfully worse results than the official numbers on some tasks.
The honest assessment: a 35B/3B MoE posting these numbers under MIT license is genuinely remarkable for builders, regardless of where it lands against the closed frontier. And because the weights are MIT-licensed rather than merely downloadable, the community can actually act on the numbers — fine-tuning, distilling, and quantizing without a legal department in the loop. The 397B’s 800GB footprint is a reminder that ‘open weights’ and ‘runnable’ are different things — the 9B and 35B are the commercially interesting releases, not the flagship.
The MIT advantage
The license is the second-biggest story here. MIT means unrestricted commercial and research use: fine-tune it, distill it, ship it in a product, never pay a royalty or ask permission. Contrast that with this summer’s other open-weights releases. Qwen3.8-Max ships under a bespoke license with revenue thresholds and a separate commercial tier for large MaaS businesses. MiniMax’s H3 community license excludes the US, EU, UK, and South Korea from its applicable territory entirely. GLM-5.2 and 5.3 are MIT, like Ornith — and that permissive pair is quietly becoming the default substrate for builders who don’t want legal review attached to every deployment.
License clarity has compounding effects. Permissive weights get more fine-tunes, more quantization builds, more integration work — the day-one GGUF, MLX, and Ollama support Ornith enjoyed is downstream of a license that lets the community do whatever it wants. In the open-weights economy, MIT is a distribution strategy.
Why it matters
Ornith-1.5 is the clearest demonstration yet that the open-weights world is experimenting with how models are trained, not just scaling up the same recipes. A self-improving training loop that works is a structural advantage: it compounds. And shipping it all under MIT — unrestricted commercial and research use — means anyone can inspect, fork, and build on the result.
Compare the licensing posture with Qwen3.8-Max’s bespoke revenue-share license to see how differently ‘open’ is being defined this summer. And if you want to serve Ornith-1.5 at scale, the inference stack just got a major upgrade — see vLLM 0.28.
Featured image: Python Software Foundation via Wikimedia Commons (GPL)
