Skip to main content

Zellige AI Labs

Frontier research, built in the open.

We study how frontier models work and publish everything we learn. Open research, open weights, no gatekeeping.

Latest from the lab

titans

Parametric Memory Cannot Replace Retrieval: Why Titans-Style Neural Memory Fails for Cross-File Symbol Resolution

We integrate Titans-style parametric memory into Qwen 3.5-9B and evaluate on CrossCodeEval for cross-file code completion. Across four architecture iterations, parametric memory produces zero measurable improvement over a memory-ablated baseline (1.0% EM vs 1.0% EM), while explicit context concatenation yields 2.5x higher exact match. Per-sample analysis traces the failure to a mechanism-task mismatch: inner-loop SGD on a low-dimensional MLP compresses lossily, but cross-file symbol resolution requires lossless recall of specific token sequences.

Read paper →

How we build

Our work follows a consistent process: early-stage research, rigorous evaluation, then production-ready systems.

01

Experiment

We explore approaches across multiple research tracks, including methods that haven't been widely studied.

02

Validate

Everything goes through blind evaluations and real-world benchmarks. Work that doesn't meet the bar is documented and published.

03

Ship

Research that meets our standards becomes available as models, APIs, or open-weight releases.

Built from our research

We're turning our research into models you can use. First releases are in development.

Mosaic Pro

Coming Soon

Our larger reasoning-focused model with thinking mode. Extended context, agentic workflows.

Mosaic Micro

Coming Soon

Fast 8B model for everyday tasks. Lightweight, efficient, surprisingly capable.

Preview chat is waitlist-only for now. Join the waitlist to get early access.

Our approach

The principles that guide our research and what we choose to invest in.

State of the Art

Our models are benchmarked against the strongest systems available. We focus on producing work that meets or exceeds current standards.

Open Source

We release model weights, publish our research, and maintain public repositories. Making our work accessible is a core part of how we operate.

Experimental

We allocate significant effort to novel architectures, training methods, and research directions that are underexplored in the broader field.

Questions we get a lot

If something isn't answered here, reach out via contact.

  • What is Zellige AI?
    An independent research lab studying how frontier AI models work and exploring new approaches to building them. We publish our findings, release open weights, and build tools from what we learn.
  • What kind of research do you do?
    We study the internal mechanics of large language models: how they represent knowledge, how information flows through layers, and what makes certain architectures work better than others. Our published work covers KV-cache injection, structural alignment in transformer attention, autograd memory optimization, and more.
  • Why publish everything?
    Too much AI research happens behind closed doors. We think the field moves faster when findings are shared openly. Every experiment we run gets documented and published, including the ones that don't work.
  • Can I use your models?
    Not yet. Our first models (Mosaic Pro and Mosaic Micro) are in development. When they ship, we plan to release open weights on Hugging Face under permissive licenses. Chat access will follow.
  • How is this different from other AI labs?
    We're smaller and research-first. No product moat, no closed API without a paper. Our architecture decisions, training choices, and evaluation results are published so anyone can learn from or build on our work.
  • How do I get involved?
    Read our research, follow us on X, or reach out if you want to collaborate. We're always interested in talking to people doing interesting work in this space.

Chat is coming soon.

Drop your email and we'll ping you the moment it goes live.

Join the waitlist →