Deep Reinforce released Ornith 1.0, and the 9B member reports 69.4 on SWE-bench Verified against 70 for Qwen3.5-35B, a model roughly four times its size (Ornith-1.0-9B model card, 2026). It pulls in 5.6GB, ships under MIT, and leaves room on a 16GB Mac. On Terminal-Bench 2.1 it does not just close the gap, it passes the 35B: 43.1 against 41.4. These are the publisher's own numbers, so treat them as claims until independent leaderboards catch up.
What is Ornith 1.0?
Ornith 1.0 is a family of open-source agentic coding models from Deep Reinforce, released in four sizes: 9B dense, 31B dense, 35B MoE, and 397B MoE. The smaller members are post-trained on top of Gemma 4 and Qwen 3.5 rather than trained from scratch, which is why a 9B can punch at this weight class.
The training method is the interesting part. Per the model card, Ornith uses reinforcement learning to generate not only solution attempts but also the scaffold that drives those attempts. Optimizing the harness and the answer together is a different bet from simply training on more code, and the benchmark spread suggests it paid off on agentic tasks specifically.
The license is MIT, described on the card as globally accessible and free from regional limitations. That is a meaningful detail in 2026, when several strong open-weight releases carry custom licenses with usage carve-outs.
How does Ornith 9B compare to bigger models?
This is the publisher's own comparison table, reproduced as published (Ornith-1.0-9B model card, 2026):
| Benchmark | Ornith 9B | Qwen3.5-9B | Qwen3.5-35B | Gemma4-12B |
|---|---|---|---|---|
| SWE-bench Verified | 69.4 | 53.2 | 70 | 44.2 |
| Terminal-Bench 2.1 (Terminus-2) | 43.1 | 21.3 | 41.4 | 21 |
| SWE-bench Pro | 42.9 | 31.3 | 44.6 | 27.6 |
| SWE-bench Multilingual | 52 | 39.7 | 60.3 | 32.5 |
| NL2Repo | 27.2 | 16.2 | 20.5 | 10.3 |
Two readings matter here. Against its own size class the result is lopsided: 69.4 against 53.2 on SWE-bench Verified, and double the score on Terminal-Bench. Against the 35B the picture is mixed, which is the honest version. Ornith 9B wins on Terminal-Bench and NL2Repo, ties on SWE-bench Verified, and loses on SWE-bench Multilingual by 8.3 points.
So the claim to carry away is not that a 9B replaces a 35B everywhere. It is that on agentic coding tasks, the ones where a model drives a terminal and iterates, the size gap stopped predicting the outcome.
Does Ornith 9B run on a 16GB Mac?
Yes, comfortably. The Ollama registry manifest puts the default build at 5.6GB of weights, verified at 200 on July 28, 2026:
ollama run ornith:9bAt 5.6GB the weights take about 35% of a 16GB machine, which leaves headroom for the KV cache, the editor, and the browser you will inevitably keep open. On a 12GB machine it still fits but the margin gets thin once context grows. The 35B sibling pulls 21.2GB and belongs on a 32GB machine or larger.
| Your Mac | Ornith build | Weights | Fit |
|---|---|---|---|
| 16GB | ornith:9b | 5.6GB | comfortable |
| 24GB | ornith:9b | 5.6GB | comfortable, long context |
| 32GB+ | ornith:35b | 21.2GB | workable |
Context is the variable that bites. Weights are a fixed cost, but the KV cache grows with every token you feed it, and agentic coding sessions feed a lot. Our RAM guide covers the sizing math, and the calculator will do it for your exact configuration.
Should you switch to it?
It depends on what you were running and why.
If you are on a 16GB Mac and your current coding model is a 7B or 8B general-purpose model, Ornith 9B is worth the download. It is a specialist, and on this evidence the specialization is real.
If you are running Qwen3.6-27B on a 24GB or 32GB machine, hold. Qwen3.6-27B scores 77.2 on SWE-bench Verified on the independent vals.ai leaderboard (vals.ai, 2026), which is 7.8 points above Ornith's vendor-reported figure, and independent beats self-reported when you are choosing what to actually use. Our local vs cloud leaderboard tracks that comparison.
One structural caveat. Ornith is a coding specialist, not a general assistant. Ask it to summarize a document or hold a conversation and you are outside its training target. That is why ModelFit's default single-answer pick stays a general-purpose model: a specialist is a second model you add, not the one you replace your daily driver with. Run npx @wecko-ai/modelfit to see what your machine gets as a general pick, then add Ornith alongside it.
For a broader look at Mac coding models at the 24GB tier, see Best Local AI Coder for Mac: Qwen3.6 vs Gemma 4.
Why does a 9B beating a 35B matter?
Because it moves the floor, and the floor is where most people actually live. A 16GB MacBook Air is the most common machine in the local AI audience, and until recently serious agentic coding meant either a 24GB minimum or a cloud API key.
The mechanism is worth naming: Ornith 9B is post-trained on Gemma 4, meaning the capability arrived through a better training recipe on an existing open base rather than a bigger model. Recipes travel faster and cheaper than parameters. That is the same pattern we tracked in How Far Behind Cloud AI Is Your Laptop?, where the lag between frontier capability and laptop-class hardware measured roughly 11 months.
The honest counterweight: one vendor-published table is not a trend. Watch for Ornith appearing on vals.ai or another independent leaderboard, and watch whether the Terminal-Bench result reproduces outside the publisher's harness.
FAQ
Can I run Ornith 1.0 9B on a 16GB Mac?
Yes. The default Ollama build is 5.6GB of weights, about 35% of a 16GB machine, which leaves room for the KV cache and your other apps. Install it with ollama run ornith:9b. The tag was registry-verified at 200 on July 28, 2026. The 35B sibling needs 21.2GB and a 32GB machine.
Is Ornith 9B better than Qwen3.5 9B for coding?
On the publisher's own benchmarks, by a wide margin: 69.4 against 53.2 on SWE-bench Verified, and 43.1 against 21.3 on Terminal-Bench 2.1. Those are Deep Reinforce's numbers on Deep Reinforce's harness, not independent results. The direction is clear enough to be worth a 5.6GB download, but wait for third-party confirmation before treating the exact figures as settled.
What license does Ornith 1.0 use?
MIT, per the model card, with no regional restrictions. That is more permissive than several recent open-weight releases, including Kimi K3, which shipped under a custom license.
Is Ornith 9B a good general-purpose model?
No, and it is not meant to be. Ornith is trained specifically for agentic coding: driving a terminal, editing repositories, iterating on failures. For chat, summarization, or general questions, a general-purpose model of the same size serves you better. Most people should run one of each rather than choosing.
How big is the Ornith 35B?
21.2GB in the default Ollama build, measured from the registry manifest on July 28, 2026. That puts it on a 32GB machine or larger. It reports 75.6 on SWE-bench Verified on its own model card, about 6 points above the 9B.
Sources
- Ornith 1.0 9B model card: https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B
- Ornith 1.0 35B model card: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B
- vals.ai SWE-Bench Verified leaderboard: https://www.vals.ai/benchmarks/swebench
- Ollama registry manifest, sizes measured July 28, 2026: https://registry.ollama.ai/v2/library/ornith/manifests/9b
Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.
The weekly local-AI refresh
New open-weight models, real Apple Silicon benchmarks, and the one model worth running on your Mac this week. Free, one email a week, unsubscribe anytime.
By subscribing you agree to our Privacy Policy and to receive the weekly email. Unsubscribe anytime.
Have questions? Reach out on X/Twitter