Will a Pixel 11 Run a Model I Brought Myself? — What the 3.5x Claim Does Not Cover
2 September 2026
Running AI on the phone itself keeps coming up. No large GPU in a desktop — the handset in your pocket writes and summarises on its own.
Google released the Pixel 11 on 20 August 2026, described as up to 3.5 times faster at on-device AI.
If I bring my own model to it, do I get that speed? And how large a model will actually fit?
This article is compiled from published information. Nothing here was measured on the device (checked 2 September 2026).
- 1. What I looked at, and what I read
- 2. What changed between the Pixel 10 and the Pixel 11
- 3. Who is the “3.5 times faster" actually about?
- 4. Is there anything at all for route A?
- 5. The trap in which file you download
- 6. How fast is it likely to be?
- 7. How large a model will fit?
- 8. Which configuration to buy for this purpose
- 9. What I could not verify here
- 10. In summary: the speed claim and your own model are different subjects
What I looked at, and what I read
“Your own model" here means a file distributed in the GGUF format, run with llama.cpp. On a phone that usually happens inside an app called Termux. It is a different thing from Gemini Nano, the model Google ships on the device — and that difference turned out to be the whole story.
I do not own a Pixel 11. What follows was cross-checked between Google’s announcement and press coverage of the specifications, third-party measurements (Geekbench and 3DMark), ARM’s own documentation on speeding up llama.cpp, and comparison articles covering speeds by handset.
What changed between the Pixel 10 and the Pixel 11
The chip moved from Tensor G5 to Tensor G6.
| Pixel 10 (G5) | Pixel 11 (G6) | |
|---|---|---|
| Process | — | 2nm |
| CPU | 8 cores (Cortex-X4 and others) | 7 cores (ARM C1 series) |
| GPU | PowerVR DXT-48-1536 | PowerVR CXTP-48-1536 |
| TPU | — | 50% more compute (Google’s figure) |
| RAM | Pro and Pro XL: 16GB at every storage size | 256GB model: 12GB |
Third-party measurements report the CPU about 18% faster single-core and about 21% faster multi-core — with one core fewer than before.
The GPU part number is stranger. The Pixel 10 has a DXT; the Pixel 11 has a CXTP. The C series precedes the D series, and that block was reportedly announced back in 2021. Measurements put peak performance only 7 to 10% higher, and after a few minutes of stress testing it fell below the previous generation. For gaming that is a step backwards. For local LLMs it was never the part I was counting on, for reasons below.
Who is the “3.5 times faster" actually about?
That figure describes Gemini Nano running on the TPU, the dedicated block Google includes for its own AI work.
A comparison article covering speeds by handset put it plainly: Tensor does not expose the NPU to third-party apps. Across six apps it tested, everything on a Pixel ran on the CPU.
A model you bring yourself never touches the TPU. There are three separate routes, and mixing them up is what makes the numbers stop adding up.
| Route | What runs it | Does the 3.5x apply? |
|---|---|---|
| A. Termux + llama.cpp (your own GGUF) | CPU. No TPU, little from the GPU | No |
| B. Google AI Edge Gallery (Google’s models) | Gemma 3n E2B (2GB) / E4B (3GB) | Yes |
| C. LiteRT / Tensor SDK (convert your model) | TPU, after the conversion work | Yes |
Is there anything at all for route A?
There is, and it is not in the GPU or the TPU. It is inside the CPU.
The ARM C1 series used by Tensor G6 includes SME2, a set of instructions for speeding up matrix multiplication. Most of what an LLM does is matrix multiplication, so this lands directly on inference. The previous Tensor G5 used Cortex-X4, a 2023 design without it — that last point is my inference from the core generation, not something Google states.
ARM’s own documentation says that when the CPU has SME2, KleidiAI replaces the generic implementation inside llama.cpp. The same document includes measurements with and without it.
| Llama-3.2-3B, Q4_0 | Reading speed | Writing speed |
|---|---|---|
| With SME2 | 12.3 tok/s | 9.1 tok/s |
| Without SME2 | 7.9 tok/s | 5.9 tok/s |
| Difference | +56% | +54% |
The handset used for that measurement is not published, so these are not Pixel 11 numbers.
The trap in which file you download
SME2 only helps two formats: Q4_0 and Q8_0.
The GGUF you are most likely to find is Q4_K_M, because it balances quality against size well. With the K and IQ families, the code silently falls back to the old path. No error, no warning.
Pick Q4_K_M in good faith and the new hardware feature never runs, with nothing on screen to tell you. Choosing Q4_0 is the whole fix — but you have to know it exists.
How fast is it likely to be?
No report of anyone running a model on a Pixel 11 turned up in my search. It had been on sale four days.
Working from existing numbers: a Pixel 9 Pro running Phi-4 Mini (3.8B) Q4_K_M on the CPU gives 10 to 18 tok/s; SME2 is worth about 1.54x in ARM’s measurement; and Geekbench 6 puts G5 to G6 at about 1.2x. Multiplied out that suggests roughly 21 to 37 tok/s.
The two foundations disagree, though. ARM measured 9.1 tok/s on a 3B model, which would make the Pixel 9 Pro faster on a larger one. Without knowing ARM’s handset there is no way to resolve it. Taking both, a 3 to 4B model lands somewhere around 10 to 35 tok/s — a range of more than double, which no amount of reading will narrow. Someone has to measure it.
How large a model will fit?
RAM bites harder than speed does.
| Model size | File at Q4_0 | 12GB phone | 16GB phone |
|---|---|---|---|
| 1B | about 0.8GB | comfortable | comfortable |
| 3B | about 1.9GB | comfortable | comfortable |
| 4B | about 2.5GB | fits | fits |
| 8B | about 4.7GB | marginal | fits |
| 12B | about 7.0GB | does not fit | tight |
Android spends RAM on the system and other apps, so a 12GB handset leaves roughly 6 to 7GB usable. Several articles agree on the same shape: 7B is about the ceiling, 3B and below is comfortable. A long conversation also consumes memory beyond the model itself, so the table assumes short exchanges.
Which configuration to buy for this purpose
| 256GB | 512GB | 1TB | |
|---|---|---|---|
| Pixel 11 | 12GB | 12GB | — |
| Pixel 11 Pro | 12GB | 16GB | 16GB |
| Pixel 11 Pro XL | 12GB | 16GB | 16GB |
| Pixel 11 Pro Fold | 16GB | 16GB | 16GB |
The Pixel 10 Pro and Pro XL carried 16GB at every storage size. On the Pixel 11 the 256GB model drops to 12GB, and the cheapest Pro XL is reported to be $100 more than last year.
| RAM | SME2 | Model size it reaches | |
|---|---|---|---|
| Pixel 10 Pro (256GB) | 16GB | no | up to 8B |
| Pixel 11 (256GB) | 12GB | yes | 3 to 4B comfortably |
| Pixel 11 (512GB and up) | 16GB | yes | up to 8B |
Newer is not automatically better here. The cheaper Pixel 11 sits below the previous generation on how much it can hold, even though SME2 should put it ahead on speed — and nothing runs at all if it does not fit. If local LLMs are the reason for the purchase, the argument for 512GB or more is not the storage. It is that the RAM becomes 16GB.
What I could not verify here
- I own neither a Pixel 11 nor a Pixel 10. All of this comes from published material
- The 50% TPU gain and the 3.5x are Google’s own figures; no third-party measurement of them turned up
- ARM does not publish which handset it measured, so those absolute numbers cannot be mapped onto the Pixel 11
- Whether llama.cpp’s GPU path runs on this PowerVR part is unknown; its Vulkan support has been described as dated
- That Tensor G5 lacks SME2 is my inference from the core generation it uses
- The effort of converting a model with LiteRT, and how fast it is afterwards, was not looked into
In summary: the speed claim and your own model are different subjects
The 3.5x figure is not about a model you bring yourself. It describes Google’s model on Google’s route.
There is still a tailwind. SME2 in the Pixel 11’s CPU is used by llama.cpp automatically, and ARM’s own measurement puts it at around 1.5x — but only for Q4_0 and Q8_0. With the common Q4_K_M the old code path returns silently.
The ceiling is set by RAM. A 12GB handset is comfortable at 3 to 4B and marginal at 8B; 16GB brings 8B into view. And nobody has published a measurement from an actual Pixel 11 yet.
The GPU went back a generation, the RAM went down and the price went up, so reviews of it as a phone are unkind. As a tool for running models on your own hardware the verdict does not match — what got worse and what got better are in different places.