Can Your Phone Run Ornith 1.0 9B? The Coding Model Everyone's Running Locally
Quick answers:
- Yes on 12GB+ phones, no on 8GB. Ornith 9B at Q4_K_M needs ~7.1GB of usable memory; 8GB phones only have ~5–6GB usable.
- Expect ~5.5–7 tokens/s (estimated) on current flagships — reading pace, not instant. Fine for chat and short snippets; long code completions take patience.
- The bigger Ornith models don't fit any phone: 35B-A3B's smallest current GGUF is 24.7GB (misses even on 24GB Androids), and 397B starts at 112.7GB.
- Check your exact phone: Ornith 9B compatibility table or the phone checker.
What is Ornith 1.0?
Ornith 1.0 is DeepReinforce's open-source (MIT) family of agentic coding models, released June 25, 2026 — 9B and 31B dense plus 35B and 397B MoE, post-trained on Qwen 3.5 and Gemma 4 bases. Its claim to fame: instead of relying on human-designed agent harnesses, it learns to write its own scaffolds during RL training. The flagship beats much larger models on SWE-Bench Verified (82.4) and Terminal-Bench 2.1.
The 9B is the one that matters for phones — and people are clearly running it: the official GGUF repo passed 2.2 million downloads in three weeks.
RAM math (9B, per quant)
| Quant | Download | Usable RAM needed* | Verdict |
|---|---|---|---|
| Q4_K_M | 5.6 GB | ~7.1 GB | 12GB+ phones ✅ · 8GB ❌ |
| Q5_K_M | 6.5 GB | ~8.0 GB | 12GB Android ✅ · 12GB iPhone ⚠️ tight |
| Q6_K | 7.4 GB | ~9.0 GB | 12GB Android ⚠️ at the line · 16GB ✅ |
| Q8_0 | 9.5 GB | ~11.2 GB | 16GB Androids only |
*At 4K context: file × 1.05 + ~0.63GB KV cache + ~0.6GB runtime, per our methodology. There's no Q3/IQ4 build in the official repo yet — when one lands, 8GB phones come into play.
Which phones run it?
- 16GB Androids (Galaxy S26 Ultra, OnePlus 15, Pixel 10 Pro…): comfortable fit, ~5.5–7 tokens/s (est.) — S26 Ultra verdict
- 12GB Androids (S24 Ultra, Pixel 9, POCO X7 Pro…): fits with ~2GB headroom, ~5.5–6 tokens/s (est.)
- 12GB iPhones (17 Pro / Pro Max / Air): fits, but tight — ~0.7GB headroom after loading; close other apps first — iPhone 17 Pro Max verdict
- 8GB phones (iPhone 16 Pro, Galaxy S24, Pixel 9a…): ❌ — ~7.1GB needed vs ~5–6GB usable. Run Qwen 3.5 4B or another 4B-class coder instead.
Is it fast enough for coding on a phone?
Honest answer: it's at reading pace, not IDE pace. Our bandwidth estimates put it at 5.5–7 tokens/s on 2025–26 flagships — below the ~8 tokens/s where generation feels smooth. That's fine for explaining code, writing short functions, or reviewing a diff on the go; a 200-line file generation will take a few minutes. Feel the difference yourself →
If you want smooth speed over top coding quality, a 4B-class model runs roughly twice as fast on the same phone.
How to run it
- Confirm your phone fits: per-phone table
- Install PocketPal (iOS/Android) or ChatterUI (Android)
- In the model hub, search
deepreinforce-ai/Ornith-1.0-9B-GGUFand download the Q4_K_M file (5.6GB, use Wi-Fi) - Set context to 4K; enable Metal on iPhone; close background apps on 12GB iPhones
FAQ
Can any phone run Ornith 35B or 397B? No. The 35B-A3B's smallest available GGUF (Q5_K_M) is 24.7GB — about 29GB with cache and runtime, more than even a 24GB gaming phone has usable. The 397B starts at 112.7GB. Both are PC/server models; on phones the 9B is the family's only option.
Ornith 9B vs Qwen 3.5 9B — which should I pick? Ornith 9B is a Qwen 3.5-based model, post-trained hard for coding. For coding tasks, pick Ornith; for general chat/vision, the stock Qwen 3.5 9B keeps multimodal support (Ornith's GGUF is text-only).
Does it work offline? Yes — after the download, everything runs on-device.
Why does my 12GB phone feel tight? iOS caps a single app at ~65% of RAM (~7.8GB usable on 12GB iPhones), and this model needs ~7.1GB — it fits, with little room. Android's 12GB tiers have ~9GB usable, which is more comfortable.
Speeds are formula estimates from our fit engine (memory bandwidth ÷ bytes per token, conservative efficiency factor), not measurements; benchmark figures are DeepReinforce's published numbers. Measured entries will replace estimates as community benchmarks land.