Ornith-1.5 releases open models trained with self-generated tasks and scaffolds
- Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE; Ornith compares those results with Claude Opus 4.8 at 85.0 and 59.0, respectively.
- The training loop jointly generates a task, a task-specific scaffold, and solution rollouts, then propagates rollout reward through all three stages during reinforcement learning.
- The task reward multiplies validity, frontier difficulty, and novelty, rewarding tasks whose scaffolds execute and verify solutions, sit near the model's capability frontier, and differ from prior tasks.
- Ornith-1.5-35B activates 3B parameters per token and reports 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified, versus 43.4 and 52.0 for Gemma 4-31B.
- The dense Ornith-1.5-9B reports 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified; the quantized Mobile version is intended for iPhone and Android deployment.
Hacker News opinions
I couldn't find how the base model was developed. Did Ornith pretrain it from scratch, or start with another open-weights model?
It looks like they post-trained Qwen 3.6. I'm interested in how much their harness and training method actually improved it.
I've been using the 35B-A3B for web scraping today. It feels on par with Qwen 3.8 27B, but runs much faster for me at Q4 rather than Q8.
I'd still like to see actual benchmarks between Ornith 35B-A3B and Qwen 3.8 27B before drawing conclusions.
I checked the release: the weights are available on the Ornith AI Hugging Face collection.
I'm hoping the results are real. A 35B-A3B MoE matters for local use on consumer hardware, and Qwen apparently is not releasing that size for its 3.8 line.
I'd rather use speculative decoding than MoE. On 2x4090 I get 90 to 100 tokens per second with strong output quality.
My quick tests say it is good at reasoning and tool use, and wastes less time thinking, but it seems a bit worse than Qwen 3.6 35B at coding. Muse Glimmer with low reasoning strength can solve code tasks faster than the Qwen MoE on my machine.
I'm running dense Qwen 3.8 27B on an OEM Spark. I stopped choosing MoE purely for speed because the dense model's higher active parameter count improves output quality.
Qwen 3.8 27B with MTP has been fast enough that I don't need 35B-A3B. I get about 70 tokens per second on an M5 Max, dropping to roughly 40 at larger contexts.
I'll compare the 9B model with Ling 3 Tiny, an 8B-A1B model, as a scout model. Ling Tiny is extremely fast but sometimes too dumb, so I hope Ornith lands in the middle.
397B is too large for my two Sparks even with NVFP4. I wish they had made it a little smaller.
The 397B model is derived from Qwen3.5-397B-A17B through post-training, so its parameter count stays exactly the same.
In five tests, it reliably claimed to be Claude without naming a version. Many open models do this because Claude is the most defined AI persona, but it is still odd.
This reads like another startup using a vague RSI claim to help close a funding round.
The page compares against Qwen 3.6 27B, but I want a comparison with Qwen 3.8 27B because the newer model is much more capable.
My own tests found Ornith-1.0-9B worse than Qwen3.5-9B, despite the published scores implying the reverse. I'll put Ornith-1.5-9B through the same tests.