Michael Limberger
Need me? Email mike@limberger.ca
Winner
🐦 Ornith
Top to bottom · draft, still filling in
Ornith won my local Skiff tests so far. Fast, clean tool use, and it behaved well inside a harness nobody trained it against. This page is the longer story. Rough draft. Details will move as I learn more.
Useful links: deep-reinforce.com · Ornith 1.0 on Hugging Face · Ornith 1.5 MLX
Who made it
DeepReinforce · Dr. Jiwei Li · San Jose
The lab
CUDA kernels, agentic RL, then Ornith
Self-scaffolding
Why it fits Skiff so well
Your 35B
MoE · multimodal · 256K context
Vendor numbers
Trust carefully · use your own harness
First three moves
Thinking budget · sampling A/B · Ornith 1.5
Who made it
DeepReinforce is a small AI startup and research lab in San Jose, California. Roughly 11 to 50 people. Site: deep-reinforce.com.
Founded by Dr. Jiwei Li. Stanford NLP PhD with Dan Jurafsky in three years. Stints at Facebook AI Research and Microsoft Research. Earlier he founded Shannon.AI (now closed) and served as Chief AI Officer at Altonomy. DeepReinforce has not disclosed its own funding.
What the lab is actually about
They did not start as a model lab. The mission is automating hard low-level software work with agentic reinforcement learning: CUDA-L1 / L2, CRINN, GrandCode, IterX, then Ornith.
Their signature trick from CUDA-L1 is contrastive reinforcement learning: compare code versions side by side and learn why some are faster. They also got burned by reward hacking, and that scar tissue shows up in Ornith's design.
What Ornith is
Released June 25, 2026 under the MIT license. Four sizes: 9B dense, 31B dense, 35B mixture of experts, and a 397B MoE flagship. Post-trained on Qwen 3.5 and Gemma 4 bases. The name comes from Greek ornis, bird.
The pack I ran is the 35B Qwen 3.5 branch, not the Gemma branch. Weights: deepreinforce-ai/Ornith-1.0-35B on Hugging Face.
The secret sauce: self-scaffolding
Normal approach: a human writes the agent harness, and you train the model to live inside it. The harness stays fixed.
Ornith's approach: the scaffold is learnable. Each RL step has the model refine a scaffold, then solve the task inside that scaffold. Reward flows to both stages. Over time, per-task strategies show up without hand-built harness design.
That is a clean explanation for why Ornith behaved well inside Skiff, a harness nobody trained against.
Anti reward-hacking
They expected cheating. Fixed trust boundary, deterministic monitor, and a frozen LLM judge. Training rollouts also ran with network access off. That is training containment, not a substitute for your own limits when you give real tools.
Architecture of the 35B
About 35B total parameters, roughly 3B active per token. 256
experts, 8 active plus 1 shared. Hybrid attention. A vision
encoder. Hugging Face tags it multimodal
(image-text-to-text). 256K context. Reasoning model:
you need a reasoning parser and a tool-call parser or the agent
loop breaks.
That vision encoder is why an mlx_vlm path can be
the right call for this pack, unlike a text-only
mlx_lm assumption.
Sampling notes
Official recipes disagree: Terminal-Bench style is temperature 1.0 / top_p 1.0; Transformers quickstart is 0.6 / 0.95 / top_k 20. My row used the quickstart values. Worth an A/B. Also: a thinking budget of 900 on a 256K reasoning model is probably too tight.
Vendor numbers, with skepticism
They report strong Terminal-Bench and SWE-bench numbers for the 35B, including beating much larger Qwen 3.5-397B on some tables. Treat those as vendor claims until third parties reproduce them. Community reception is mixed. My own Skiff harness is there for a reason.
First three things to do
Raise thinking budget
Try something like 8000 instead of 900 on the ornith row.
A/B sampling
0.6 / 0.95 / 20 versus 1.0 / 1.0 on the same tasks.
Add Ornith 1.5
Official MLX build ornith-ai/Ornith-1.5-35B-A3B-MLX as a separate catalog row, not a replace.