AI Models for Pixel 9 — What runs on 12GB
Specs checked against manufacturer and public documentation on .
What runs on the Pixel 9
All 48 models at their recommended quant, on the 12GB configuration. Tap a model for the full report.
| Model | Params | Quant | Needs | Speed | Verdict | |
|---|---|---|---|---|---|---|
| S | Qwen3 0.6B | 0.6B | Q8_0 | 1.3 GB | ~51.2 tokens/s | ✓ Runs great |
| S | Qwen3 1.7B | 1.7B | Q8_0 | 2.6 GB | ~17.1 tokens/s | ✓ Runs great |
| S | Llama 3.2 1B | 1.2B | Q4_K_M | 1.5 GB | ~38.4 tokens/s | ✓ Runs great |
| S | Llama 3.2 3B | 3.2B | Q4_K_M | 2.9 GB | ~15.4 tokens/s | ✓ Runs great |
| S | Gemma 3 1B | 1B | Q4_K_M | 1.5 GB | ~38.4 tokens/s | ✓ Runs great |
| S | DeepSeek R1 Distill 1.5B | 1.8B | Q4_K_M | 1.9 GB | ~27.9 tokens/s | ✓ Runs great |
| S | SmolLM2 1.7B | 1.7B | Q4_K_M | 1.9 GB | ~27.9 tokens/s | ✓ Runs great |
| S | SmolLM3 3B | 3.1B | Q4_K_M | 2.8 GB | ~16.2 tokens/s | ✓ Runs great |
| S | Qwen 3.5 2B | 2B | Q4_K_M | 2.1 GB | ~23.6 tokens/s | ✓ Runs great |
| S | Ternary Bonsai 4B | 4B | PQ2_0 | 2 GB | ~27.9 tokens/s | ✓ Runs great |
| S | Ternary Bonsai 1.7B | 1.7B | PQ2_0 | 1.2 GB | ~61.5 tokens/s | ✓ Runs great |
| S | LFM2.5 8B-A1B | 8B | Q4_K_M | 6.6 GB | ~47.3 tokens/s | ✓ Runs great |
| S | Ministral 3 3B | 3B | Q4_K_M | 3 GB | ~14.6 tokens/s | ✓ Runs great |
| S | Ternary Bonsai 8B | 8B | PQ2_0 | 3.5 GB | ~14 tokens/s | ✓ Runs great |
| S | Qwen3 4B | 4B | Q4_K_M | 3.5 GB | ~12.3 tokens/s | ✓ Runs great |
| S | Gemma 3 4B | 4.3B | Q4_K_M | 3.5 GB | ~12.3 tokens/s | ✓ Runs great |
| S | Phi-4 Mini 3.8B | 3.8B | Q4_K_M | 3.5 GB | ~12.3 tokens/s | ✓ Runs great |
| S | Qwen 3.5 4B | 4B | Q4_K_M | 3.7 GB | ~11.4 tokens/s | ✓ Runs great |
| S | Nemotron 3 Nano 4B | 4B | Q4_K_M | 3.8 GB | ~11 tokens/s | ✓ Runs great |
| S | Gemma 4 E2B | 2B | Q4_K_M | 4 GB | ~9.9 tokens/s | ✓ Runs great |
| A | Bonsai 27B (1-bit) | 27B | Q1_0 | 6.5 GB | ~8.1 tokens/s | ✓ Runs great |
| A | Mistral 7B v0.3 | 7.2B | Q4_K_M | 5.7 GB | ~7 tokens/s | ! Runs, barely |
| A | Llama 3.1 8B | 8B | Q4_K_M | 6.3 GB | ~6.3 tokens/s | ! Runs, barely |
| A | Ministral 8B | 8B | Q4_K_M | 6.3 GB | ~6.3 tokens/s | ! Runs, barely |
| A | DeepSeek R1 Distill 7B | 7.6B | Q4_K_M | 6.1 GB | ~6.5 tokens/s | ! Runs, barely |
| A | Qwen3 8B | 8.2B | Q4_K_M | 6.4 GB | ~6.1 tokens/s | ! Runs, barely |
| A | Gemma 4 E4B | 4B | Q4_K_M | 6.1 GB | ~6.1 tokens/s | ! Runs, barely |
| A | Ministral 3 8B | 8B | Q4_K_M | 6.6 GB | ~5.9 tokens/s | ! Runs, barely |
| A | Qwen 3.5 9B | 9B | Q4_K_M | 7.2 GB | ~5.4 tokens/s | ! Runs, barely |
| C | Gemma 4 12B | 12B | Q4_0 | 8.8 GB | ~4.4 tokens/s | ! Runs, barely |
| F | Gemma 3 12B | 12.2B | Q4_K_M | 9.1 GB | — | ✕ Won't fit |
| F | Ternary Bonsai 27B | 27B | PQ2_0 | 10.1 GB | — | ✕ Won't fit |
| F | Ministral 3 14B | 14B | Q4_K_M | 10.2 GB | — | ✕ Won't fit |
| F | Qwen3 14B | 14.8B | Q4_K_M | 11.1 GB | — | ✕ Won't fit |
| F | Phi-4 14B | 14.7B | Q4_K_M | 11.2 GB | — | ✕ Won't fit |
| F | GPT-OSS 20B | 21B | MXFP4 | 14.8 GB | — | ✕ Won't fit |
| F | Gemma 4 26B-A4B | 26B | Q4_0 | 17.5 GB | — | ✕ Won't fit |
| F | Qwen3 30B A3B | 30.5B | Q4_K_M | 22.3 GB | — | ✕ Won't fit |
| F | Qwen 3.5 27B | 27B | Q4_K_M | 20 GB | — | ✕ Won't fit |
| F | Qwen 3.5 35B-A3B | 35B | Q4_K_M | 26.2 GB | — | ✕ Won't fit |
| F | Qwen 3.6 27B | 27B | Q4_K_M | 20.1 GB | — | ✕ Won't fit |
| F | Qwen 3.6 35B-A3B | 35B | Q4_K_M | 26.3 GB | — | ✕ Won't fit |
| F | Nemotron 3 Nano 30B-A3B | 30B | Q4_K_M | 28.5 GB | — | ✕ Won't fit |
| F | Llama 3.3 70B | 70B | Q4_K_M | 50.1 GB | — | ✕ Won't fit |
| F | Qwen3 32B | 32.8B | Q4_K_M | 23.7 GB | — | ✕ Won't fit |
| F | Gemma 4 31B | 31B | Q4_K_M | 22 GB | — | ✕ Won't fit |
| F | Hunyuan 3 (Hy3) | 298.8B | Q4_K_M | 212.8 GB | — | ✕ Won't fit |
| F | Inkling | 952.4B | Q8_0 | 966.9 GB | — | ✕ Won't fit |
~ = bandwidth-based estimate · ✓ = measured on real hardware
Best model by use case
Top everyday assistant & writing pick here — ~51.2 tokens/s at Q8_0, using 1.3 of ~9GB.
Top code completion & explain-this pick here — ~12.3 tokens/s at Q4_K_M, using 3.5 of ~9GB.
Top math & step-by-step thinking pick here — ~27.9 tokens/s at Q4_K_M, using 1.9 of ~9GB.
FAQ
What is the biggest AI model the Pixel 9 can run?
Bonsai 27B (1-bit) (27B parameters) at Q1_0 — it needs 6.5GB of the ~9GB usable on the 12GB Pixel 9 and runs at ~8.1 tokens/s.
How much of the Pixel 9's 12GB RAM can AI models actually use?
About 9GB. Android keeps roughly 2–4GB for the system and resident apps, so of the 12GB about 9GB is actually available to a model.
Can the Pixel 9 run Llama 3.1 8B?
Yes — at Q4_K_M it needs 6.3GB of the ~9GB usable and runs at ~6.3 tokens/s.
How fast is local AI on the Pixel 9?
The Google Tensor G4 has 68.3GB/s of memory bandwidth, which is what decode speed scales with. Small models like Ternary Bonsai 1.7B reach ~61.5 tokens/s; larger 7–14B models land in the single digits. Anything above ~8 tokens/s feels smooth for chat.
Which quantization should I use on the Pixel 9?
Q4_K_M is the size/quality sweet spot for most models. For example, Qwen3 0.6B at Q8_0 takes 1.3GB of memory here. Only drop to Q3 or IQ4 if a model just misses fitting; Q8 rarely pays off on 12GB of RAM.
Is 12GB of RAM enough for local AI?
30 of the 48 models we track fit on the Pixel 9 — 21 run great and 9 run with compromises. 18 models (mostly 12B+) don't fit at their recommended quant.