Experimental Machines

Open-weight models compiled for phones, with published recipes, checksums, smoke tests, and benchmark evidence.

experimentalmachines.org · experimentalintelligence.org · GitHub · OpenWeights on Google Play

Model catalog

This organization has 18 public model repositories. Collections are the catalog and recommendation surface; they group deployment exports by family and keep the organization page focused on artifacts that run in OpenWeights.

Collection Contents
LFM2.5 for ExecuTorch LFM2.5 1.2B and 2.6B exports.
LFM2.5 Abliterated for ExecuTorch The 1.2B and 2.6B heretic exports.
Qwen3 for ExecuTorch Qwen3 exports.
Qwen2.5 for ExecuTorch Qwen2.5 exports.
Llama 3.2 for ExecuTorch Llama 3.2 exports.
SmolLM2 for ExecuTorch SmolLM2 exports.
QwenGrad for ExecuTorch The evaluated deployment build from the OpenGrad study.

Every compiled export card records its source revision, quantization recipe, fixed context window, file checksum, and host-runner smoke-test result. Benchmark results are reported separately, so a runnable export is never represented as a behavioral recommendation without device evidence.

The LFM2.5 entries use the corrected prompt-state recipe and GPTQ-solved int4 weights. The heretic variants are published only as their validated ExecuTorch deployment exports; there are no separate bare heretic source-checkpoint repositories in this organization.

Research artifact

QwenGrad-DPO is intentionally outside the collections. It is the standard Transformers-format checkpoint from the OpenGrad research study; its paired phone deployment export is in the QwenGrad collection.

How exports are made

executorch-model-exporter builds supported models on GitHub-hosted runners, smoke-tests them with the same runner used by the app, and publishes the .pte programs, reports, and app-readable configuration here. The supported delegates are XNNPACK, Qualcomm QNN, MediaTek NeuroPilot, and Vulkan where the model family has a validated recipe.

Phone measurements and their raw results are published at latency, context window, and reruns.

Contact: alpha@experimentalmachines.org ยท arjhine@experimentalmachines.org