WaveletLM – wavelet-based, attention-free model with O(n log n) scaling
WaveletLM is a wavelet-based, attention-free architecture that replaces self-attention with learned lifting wavelet decomposition, a Fast Walsh-Hadamard Transform, per-scale gated spectral mixing with SwiGLU activation, an inverse FWHT, and wavelet reconstruction. Combined with expanded MLPs and sparse product-key memory, this yields a model with O(n log n) scaling in sequence length. With 23.8 PPL on WikiText-103, WaveletLM beats both GPT-2 Medium, which was trained on 80× more data, and Transformer-XL Standard, which uses recurrence to extend its effective context. It is undertrained and…
What it does
In the maker’s words, at launch
WaveletLM is a wavelet-based, attention-free architecture that replaces self-attention with learned lifting wavelet decomposition, a Fast Walsh-Hadamard Transform, per-scale gated spectral mixing with SwiGLU activation, an inverse FWHT, and wavelet reconstruction. Combined with expanded MLPs and sparse product-key memory, this yields a model with O(n log n) scaling in sequence length. With 23.8 PPL on WikiText-103, WaveletLM beats both GPT-2 Medium, which was trained on 80× more data, and Transformer-XL Standard, which uses recurrence to extend its effective context. It is undertrained and underregularized due to budget constraints, so there is much room for development and improvement. I invite anyone who is curious to examine the model, test it out, and extend its capabilities further. All code and weights are fully open source, and a PG-19 run will be completed in 2-3 days. Generations can be done in 4-5 GB VRAM at 28.8 tokens/second, and the model is trainable in 16.25 hours with 20 GB of VRAM, both on a 5090. README for comparison tables, instructions, logs, and future plans: https://github.com/ramongougis/WaveletLM Weights: https://huggingface.co/ragou19/WaveletLM/tree/main Generations: https://github.com/ramongougis/WaveletLM/blob/main/logs/wiki... The following samples were chosen for coherence, not factual accuracy. Factuality will require scaling and downstream techniques such as RAG and instruction tuning. > The history of the city is reflected in its architecture, which includes the historic Old Town and New Castle County Courthouse Square Historic District. The building was designed by John H. Stevens, who also designed the Albany-Fulton Celebration in 1906 and built a steel-hulled shipyard on the lake shore. > The album was released on August 25, 2007 by Sony Music Entertainment and features several songs from the record including "Never Say Die", "The Show", "Don't Cry for Me Argentina" and a cover of "I Can Only Imagine (But You Are Not Alone)". > The species was first described by Swedish zoologist Carl Linnaeus in 1758 as Agaricus adustus. The genus name is derived from the Latin words perma "to tie", and pous ("like") means "with a large head". In 1821, French mycologists Jean-Baptiste de Lacaille placed it in section Cricetae of the order Carnivora. He later renamed it Spongiforma punctata after the Greek kribensis.
Does the same job
all alternatives →
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, April 2026
the whole month →
- AG
Thought the resources for GPU arch were lacking, so here we are
Life & fun · Apr 2026 · jaso1024.com
- IB
Built a ~9M param LLM from scratch to understand how they actually work. Vanilla transformer, 60K synthetic conversations, ~130 lines of PyTorch. Trains in 5 min on a free Colab T4. The fish thinks the meaning of life is food. Fork it and swap the personality for your own character.
AI · Apr 2026 · github.com
- BC
Life & fun · Apr 2026 · sam-burns.com
- IB
With social media and now AI, its important to keep the indie web alive. There are many people who write frequently. Blogosphere tries to highlight them by fetching the recent posts from personal blogs across many categories. There are two versions: Minimal (HN-inspired, fast, static): https://text.blogosphere.app/ Non-minimal: https://blogosphere.app/ If you don't find your blog (or your favorite ones), please add them. I will review and approve it.
AI · Apr 2026 · text.blogosphere.app