AmSach/maple
Can I trust it?
Be careful with this one.
Warnings Check
Very new: it was created 9 days ago. Wait a little before installing it.
Stars from real accounts No data
We haven't checked its stars for fakes yet.
Active OK
Last change 2 days ago.
Licence Careful
No licence: legally you can't reuse the code, unless the README says otherwise.
Quick checks from public data. They help, but they are no guarantee.
What it is
Research project that prunes low-activation experts from Mixture-of-Experts LLMs using routing statistics from a calibration corpus, with a reference Gemma 4 26B-A4B model reduced to about 4.75 GB for GPU execution. Provides profiling and build scripts for porting the method to…
Who it's for
ML researchers and developers compressing MoE language models for local or edge GPU deployment
How to use it
Ask your AI to set it up
Paste this into Claude, ChatGPT or your coding assistant.
Help me install and set up AmSach/maple in my project. Follow its README: https://github.com/AmSach/maple. Explain each step and check it is safe before running anything.
Star curve
On Hacker News
- Show HN: Maple – Sub-2-BPW Moe Pruning to Run Gemma 4 26B at 45 TPS on a 6GB GPU21 points · 5 comments · 8 Nov 2026
Get alerts for repos like this
We'll email you when something in Models & inference starts rising. Free, once a week, only what matches.
What arrives: one email on Mondays with the repos that match, never more. See last week's. Privacy
For maintainers
Add this badge to your README. It shows the repo's place on the radar, updates every hour and links back to this page.
[](https://repoalerts.com/repo/AmSach/maple?ref=badge)Similar repos
- freriadoc/Avalon
A fork of llama.cpp tuned for running the Qwen3.8-Flash-Next 125B MoE model on a single consumer AMD GPU with experts offloaded to system RAM. It uses an expert cache in VRAM, CPU computation of missi
- prabanta-dev/janas
A C inference engine for running large mixture-of-experts language models on ordinary computers, with or without a GPU. Weights stay on disk, and the experts each token needs are streamed from an NVMe
- mw00/project-maya
Runs GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 locally on one or more NVIDIA or AMD GPUs, with an engine, server, browser dashboard, quantization tools and OpenAI- and Anthropic-compatible API. Exper
- FlashML-org/FreeToken
Edge-native Mixture-of-Experts serving engine for running large open-weight MoE models on consumer hardware. It combines CPU and GPU execution, expert caching and KV cache management, and exposes Anth
- RustingStudio/RustingStrata
Pure-Rust inference engine for Qwen3.5/3.6 MoE GGUF models, with CUDA kernels compiled at run time via NVRTC and a Vulkan backend. Runs experts that do not fit in VRAM on the CPU alongside the GPU and
- YeonwooSung/llm-architecture-from-scratch
A collection of from-scratch PyTorch implementations of modern LLM architectures, including GPT-2 XL, Llama variants, DeepSeek V3, Qwen3, and Gemma 3. A comparison table lists each model's normalizati