Independent measurements of AI hardware, small models trained in the open, and tools to run open weights on a phone. We publish the measurements, methods, and raw results behind our claims.
| Site | Focus |
|---|---|
| experimentalmachines.org | AI accelerator benchmarks from datacenter to phone scale. Source: experimentalmachines.org. |
| experimentalintelligence.org | Small models trained from scratch, with weights, code, and training logs. Source: experimentalintelligence.org. |
OpenWeights is an Android app for open-weight language models. It searches Hugging Face in the app, estimates whether a model fits before download, and runs locally without an account, cloud service, or telemetry. It supports GGUF through llama.cpp and compiled .pte programs through ExecuTorch. It is available on Google Play under Apache 2.0.
The phone measurements behind the app include latency on five chips, the exported-window study, and reruns.
The Experimental Machines Hugging Face organization has 18 public model repositories. The catalog is organized through collections; use them as the model navigation surface.
| Collection | Contents |
|---|---|
| LFM2.5 for ExecuTorch | LFM2.5 1.2B and 2.6B ExecuTorch exports. |
| LFM2.5 Abliterated for ExecuTorch | The 1.2B and 2.6B heretic exports, with 2k through 32k windows. |
| Qwen3 for ExecuTorch | Qwen3 deployment exports. |
| Qwen2.5 for ExecuTorch | Qwen2.5 deployment exports. |
| Llama 3.2 for ExecuTorch | Llama 3.2 deployment exports. |
| SmolLM2 for ExecuTorch | SmolLM2 deployment exports. |
| QwenGrad for ExecuTorch | The evaluated ExecuTorch deployment build for the OpenGrad study. |
Each export card identifies the source revision, quantization recipe, context window, file checksum, and smoke-test result. The compiled LFM2.5 exports use the fixed prompt-state recipe and GPTQ-solved int4 weights; benchmark reports distinguish model behavior from export validation.
QwenGrad-DPO is intentionally outside the collections because it is the standard Transformers-format research checkpoint from the OpenGrad study, rather than an app deployment artifact. Its paired ExecuTorch program is in the QwenGrad collection.
OpenGrad is a controlled post-training study of small open-weight models. It publishes pre-registered evaluation gates, per-example records, negative findings, weights, code, and training logs. The frozen evidence is tagged study-001, with reports at opengrad.arjhinety.com.
executorch-model-exporter exports supported open-weight models on GitHub-hosted runners, smoke-tests each .pte with the same runner used by the app, and publishes the artifacts and reports to Hugging Face. The exporter supports XNNPACK, Qualcomm QNN, MediaTek NeuroPilot, and Vulkan where a family has a validated recipe.
alpha@experimentalmachines.org · arjhine@experimentalmachines.org
