Every desk has a default local model. Ours is not Gemma 4, not Qwen, not Nemotron — it is a 9-billion-parameter model from Ornith AI, the lab associated with DeepReinforce, trained by a loop that generates its own tasks and builds its own harness. Here is why.
October 21, 2026 · 7 min read
Most open-model releases are a bigger number on a familiar chart. Ornith-1.5, released by Ornith AI on August 20, 2026, is interesting for a different reason: the training method is the product. The family — a 397B MoE flagship, a 35B-A3B MoE, and a dense 9B with a mobile variant — is MIT-licensed, ships with no first-party API and no token price, and was quite deliberately released as a self-host-or-run-local artifact. For anyone building a sovereign local stack, that combination alone earns attention. The recipe earns the rest.
Standard RL post-training runs on fixed, human-curated task sets inside hand-designed harnesses. Ornith-1.5 closes the loop: the training process jointly optimizes three things at once — the tasks it generates for itself, the scaffold (the harness of tools and prompts) it builds to attack each task, and the solution rollouts, with GRPO reward propagating across all three stages. It extends Ornith-1.0, which was itself built on top of Qwen 3.5 and Gemma 4 bases with continued pretraining, mid-training, and post-training — the 1.0 loop optimized scaffolds and rollouts; 1.5 adds task generation, so the curriculum itself is learned.
One honesty note, because it matters: "self-improving" describes the training loop, not a runtime feature. The checkpoint you download does not keep improving on your machine. What you get is a policy that was produced by that loop — and, for researchers, a published recipe for running the loop yourself. That recipe is why this desk studies Ornith as a reference design for our own local training work: a Qwen/Gemma-hybrid that builds its own harness is a blueprint, not just a download.
The flagship numbers are eye-catching — the 35B-A3B's model card posts 86 on SWE-bench Verified, 65.1 on SWE-bench Pro, 86.1 on Terminal-Bench 2.1, and 71.2 on Tool Decathlon, and the 397B claims a Terminal-Bench tie with Claude Opus 4.8 (86.1 vs 85.0, self-reported). But the model we actually reach for is the dense 9B. Its card reports 70.6 on SWE-bench Verified, 47.0 on Terminal-Bench 2.1 in a Claude Code harness, and 86.4 on GPQA Diamond — numbers in the class of models three times its size. It carries the family-wide 262,144-token context, thinks in <think> traces with Qwen3-compatible tool-call parsing, and ships in FP8, GGUF, MLX, and NVFP4 builds, with vLLM, llama.cpp, and Ollama support.
The hardware math is the sovereign part. At BF16 the 9B is about 19GB; at Q4 it needs roughly 6GB of VRAM, and an independent RTX 5090 test (KGP Talkie) measured the quantized 9B holding its full 262K window in 15.1GB — one consumer card, full context, no cloud. The same test found Ornith reaching answers about five times faster than Gemma 4-26B-A4B on hard questions, because it re-checks itself less; a separate hands-on comparison found its JSON output the tightest of the local field — 16 tokens where Qwen used 21 and Gemma 34 — the behavior of a model built for tool-calling pipelines rather than chat. The card has crossed 850,000 downloads on Hugging Face. Small, MIT, local, and pointed at agents: that is the shape of a default.
All three are strong, and the Open Source Watch tracks them all. Gemma 4 is lighter on VRAM at the margin, faster on long prompts, and better under many-user load. Qwen's line is broader, its MTP head is better, and at least one careful hands-on (daily.dev, on an M5 Max) found Qwen 3.8 27B beating Ornith in practical coding and tool reliability. NVIDIA's Nemotron line is excellent, openly licensed, and sized cleverly for single GPUs. The case for Ornith as the default is not that it wins every test — it is that its center of gravity is exactly the local-agent job: agentic coding and tool use at small scale, a license with nothing to lawyer over, no API dependency, no vendor account, no phone-home. Sovereignty is a property of the whole arrangement, not a benchmark row.
Receipts discipline applies to favorites too. The headline scores are the lab's own; independent verification is still catching up. KGP Talkie's hardware tests confirmed the context and speed claims but could not confirm the "wide margins" on agentic coding with their small suite; one HN tester's run of the 35B came in meaningfully below the official numbers; BenchLM's composite ranks Qwen3.6-35B-A3B ahead of Ornith's 35B (43.36 vs 34.11, on overlapping ranges and "Estimated" evidence). The <think> habit costs roughly 3.5× the tokens of a direct answer — worth it when correctness matters, waste when it doesn't — and the card's recommended sampling settings have drawn skepticism in testing. None of that dethrones it here. It defines what to verify on your own workloads before you standardize — which is, not coincidentally, the entire point of the Open Source Watch, where Ornith now sits on the radar next to the giants it is displacing on our machines.
Sources: Ornith-1.5 model cards and release coverage — orcarouter.ai, explainx.ai, AlphaSignal; independent tests — KGP Talkie, BenchLM.