{"name":"maple","repo":"AmSach/maple","url":"https://midorreal.com/project/amsach-maple","repository":"https://github.com/AmSach/maple","description":"MAPLE: Manifold-Activated Pruning of Latent Experts for Sub-2-BPW MoE LLMs","facts":{"language":{"value":"Python","source":"https://github.com/AmSach/maple","checked":"21 Nov 2026"},"releases":{"value":"0","source":"https://github.com/AmSach/maple/releases","checked":"21 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/AmSach/maple/graphs/contributors","checked":"21 Nov 2026"},"archived":{"value":"no","source":"https://github.com/AmSach/maple","checked":"21 Nov 2026"}},"stars":65.0,"writeup":{"what_it_is":"MAPLE is a method and a pruned model for Mixture-of-Experts LLMs. It profiles expert activations on a calibration corpus, scores experts by routing frequency and activation energy, and removes the lowest-scoring experts. The repository provides a pruned Gemma 4 26B-A4B GGUF file of 4.75 GB built from 48 of its 128 experts.","audience":"Developers running large MoE language models on limited GPU memory","claims":[{"kind":"specific","claim":"Compresses Gemma 4 26B-A4B from 9.92 GB to 4.75 GB","excerpt":"Compressing **Gemma 4 26B-A4B** from **9.92 GB to 4.75 GB**","status":"not_checked"},{"kind":"specific","claim":"Runs fully in VRAM on a 6 GB laptop GPU at 45.4 tokens per second","excerpt":"enabling **100% In-VRAM GPU Execution** on 6 GB laptop GPUs with **45.4 Tokens/Second** throughput","status":"not_checked"},{"kind":"specific","claim":"Retains 48 of 128 experts and removes 80 per layer","excerpt":"We prune the 80 dormant experts and preserve the 48 active ones","status":"not_checked"},{"kind":"specific","claim":"Scoring experts captures 79.88% of active forward-pass activation mass","excerpt":"capturing **79.88% of active forward-pass activation mass**","status":"not_checked"},{"kind":"specific","claim":"Pruned model retains unit-test execution on Python tasks within its target domain","excerpt":"the model retains 100% functional unit-test execution when given standard token generation headroom","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-09"},"why_now":null,"owner_supplied":null,"score":{"verdict":"too_early","hype":90,"reality":51,"gap":39,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":{"estimate":true,"lean":"too_close","gap_low":16,"gap_high":57,"rule":"early-2.0"},"changes":[{"date":"2026-11-07T17:32:34+00:00","text":"Repository created","source":"https://github.com/AmSach/maple"}],"cite":"maple is a GitHub project at https://github.com/AmSach/maple written mainly in Python, described there as \"MAPLE: Manifold-Activated Pruning of Latent Experts for Sub-2-BPW MoE LLMs\".","attribution":"Data from Mid or Real, https://midorreal.com/project/amsach-maple, checked 22 November 2026. Attribution with a link is required.","still_real":null}