We build apps, models and runtime fixes that bring open models to phones.
What we measured, on which hardware, and how. Every study links to its committed results.
1%
of its writing speed left: the app's engine at 4 threads, with six of its eight cores taken out of use by hand
Newest: 3 to 11 November 2026
When a Snapdragon 8 Elite Gen 5 phone overheats, it stops using six of its eight cores, and llama.cpp, the engine that runs the model, nearly stops writing the reply if it uses more than two threads. Why, the fix OpenWeights now carries, and whether ordinary chat gets that hot.
9 of 9
models were fastest with NPU prefill
OpenWeights, 2 to 10 November 2026
Prefill and decode on CPU, GPU and NPU
A model reads the prompt (prefill), then writes the reply a token at a time (decode). Nine models on a Snapdragon 8 Elite Gen 5, with each job on the CPU, the GPU or the NPU, the chip built for AI.
42%
of 24 real tasks with Qwen3.6 on an RTX 4090, against 92% on Copilot Pro+ Auto
GGUF, October 2026
When is local Copilot worth it?
The same Copilot CLI agent on the same real coding tasks: downloadable models running locally on an RTX 4090 and a MacBook Air M5, against hosted Copilot.
8 agents
in a simulated coding workload, at a 4.9 s wait per turn on vLLM with no errors. At 16, waits reach 16.5 s
Hardware, 23 and 24 September 2026
How many coding agents can one RTX 4090 serve?
Qwen3.8-27B on one rented RTX 4090, through five engine setups, rising load and real coding agents.
Where we take what we measured. Every talk links to its listing and the evidence behind it.
PyTorch Day Japan
27×
faster prefill on a second turn: Qwen3-1.7B on ExecuTorch, reusing 1,192 tokens
Tokyo, 10 December 2026
When the KV Cache Cannot Roll Back
How a phone skips re-reading a 2,050-token agent prompt, built twice because the app's two runtimes have different cache constraints.
Alpha Romer Coma and Arjhine Ty
PyTorch Day Korea
1 in 5
replies from the compiled 1.2B cited a search it never made, on our first test set. As a GGUF, none of 640
Seoul, 21 November 2026
Whether a small model on a phone searches when it must, on 160 public questions, four phones and two builds of the same weights.
Xynil Jhed Lacap
GitHub Universe
24 tasks
of real work on ten setups. Each bar is tasks finished; the bright ones ran locally
San Francisco, 28 October 2026
When is local Copilot worth it?
The same Copilot CLI agent on the same real tasks: open-weight models you serve yourself against hosted Copilot.
Alpha Romer Coma
PyTorch Conference Europe
Compiler Tax
what surprises developers moving PyTorch workloads to ASICs
Paris, 8 April 2026
De-mystifying PyTorch for ASICs
A code-first check on moving PyTorch from GPUs to Google TPUs and AWS Trainium, and when it is worth doing.
Alpha Romer Coma
OpenWeightsOn Google Play
Search Hugging Face, see whether a model fits your phone, and run it there. No account.
ExecuTorch buildGGUF in llama.cpp
ExecuServeOn Google Play
A local AI server on your Android phone, behind the OpenAI and Anthropic APIs.
Call it with the OpenAI SDK you already use: set OPENAI_BASE_URL to the address the app shows, and OPENAI_API_KEY to its key.
from openai import OpenAI
client = OpenAI() # reads OPENAI_BASE_URL and OPENAI_API_KEY
reply = client.chat.completions.create(
model="qwen3-1.7b",
messages=[{"role": "user", "content": "Hello"}],
)Call it with the Anthropic SDK you already use: set ANTHROPIC_BASE_URL to the address the app shows, without the /v1, and ANTHROPIC_API_KEY to its key.
from anthropic import Anthropic
client = Anthropic() # reads ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY
reply = client.messages.create(
model="qwen3-1.7b",
max_tokens=200,
messages=[{"role": "user", "content": "Hello"}],
)Time to the first token, ~700-token prompt
ExecuPackOn Hugging Face
Open models built for phone CPUs, GPUs, and Qualcomm and MediaTek NPUs.
Scroll sideways for all four targets.
| Model | CPUXNNPACK | GPUVulkan | Qualcomm NPUSM8750 | MediaTek NPUMT6991 |
|---|---|---|---|---|
| Qwen3 1.7BAnswering in the ExecuServe chat above | Published, windows 2k–32k | Published, windows 2k–32k | Published, windows 2k–16k | Published, windows 2k–8k |
| LFM2.5 1.2B InstructRecommended in OpenWeights | Published, windows 2k–32k | Not published | Not published | Published, windows 512–8k |
| Llama 3.2 1B Instruct | Published, windows 2k–32k | Published, windows 2k–32k | Published, windows 2k–16k | Published, windows 2k–16k |
| SmolLM2 135M Instruct | Published, windows 2k–32k | Published, windows 2k–32k | Published, windows 2k–8k | Published, windows 2k–16k |
| Qwen2.5 1.5B Instruct | Published, windows 2k–32k | Published, windows 2k–32k | Not published | Not published |
| Qwen3 4B | Published, windows 2k–32k | Published, windows 2k–32k | Not published | Published, windows 2k–4k |
MediaTek position encoding
MediaTek's stock export scripts built Qwen and Llama position tables at the wrong base. After our patch, three models matched the Hugging Face reference in a 301-token CPU test before quantization.
Finding 37 in the research log
Embeddings that differed by machine
The same model built on two machines decoded different text. The cause was a platform-dependent search in embedding quantization; new exports use a deterministic method.
Finding 27 in the research log
Our apps turn up bugs in the runtime underneath. We report them to the ExecuTorch team, and fix them where we can.
Fixed
On Android, a default temperature of 0.8 overrode the one the app set.
Our report; the maintainers' fix merged September 30, 2026.
Open
The exporter and the runner disagreed by one token about the prompt limit, so some prompts failed to prefill.
Our report and fix, opened October 10, 2026.
Alpha Romer Coma
Co-FounderLinkedIn, Alpha Romer Coma
John Andrew Balbarosa
Co-FounderLinkedIn, John Andrew Balbarosa
Arjhine Ty
Co-FounderLinkedIn, Arjhine Ty
Xynil Jhed Lacap
Co-FounderLinkedIn, Xynil Jhed Lacap
JC Diamante
Core ContributorLinkedIn, JC Diamante
Jules Gregory Agustin
Core ContributorLinkedIn, Jules Gregory Agustin
Ship something with us and your name goes here.
Work with us
Everyone has a theory about what a phone can run. Bring yours, and we'll put it on real hardware and find out.
You have a hunch.
We'll test it on real phones and show the numbers, as we did for which processor should prefill and which should decode.
See the studyYou have a model.
We'll build it for phone CPUs, GPUs and Qualcomm and MediaTek NPUs, and tell you which context windows fit.
See the buildsYou have an Android app.
We'll run the model on the phone behind the OpenAI and Anthropic APIs, so your app calls it like a cloud model.
See ExecuServeYou want to build.
Our apps and build tools are open source. Ship something with us and your name goes on this page.
See the codeThree lines is enough: what you want to test, on what, and why. It goes straight to Alpha, not a form.
alpha@experimentalmachines.org