AI & ML interests
None defined yet.
Recent Activity
Lowdown Labs
AI you can feel better using.
The AI brand for people who give a damn. We build frontier models the frontier way: fair to creatives, kind to the planet, and radically transparent.
"AI for good, for real" is not just a silly motto to us - it really matters, and it's why we are here.
What we are shipping
We build small, maximally efficient models that run on the hardware you already own, no coal rolling mega GPU data centers required.
FEVER
Fourier Encoder for Visual Embedding and Retrieval
The first truly cross resolution text to image and image to text retrieval model. No resizing tax, no quality cliff - you can embed once and query off a thumbnail.
Built off of our FELA architecture, FEVER ships both as a model and as a bundled retrieval system, available to self deploy on AWS 3rd Party and via API. We maintain task specific variants as well for certain domains.
| Product | Description | Use Case / Benefit | Key Advantage |
|---|---|---|---|
| FEVER Base | General purpose FEVER embedding and retrieval model. | Index photos, supports text -> image and image -> image retrieval. | Probably the only model that can index billions of photos in reasonable time. |
| FEVER Geo | Geospatially tuned FEVER embedding and retrieval model. | Rapid indexing and search of geospatial satellite images. | Works best with low resolution satellite imagery. |
| FEVER MRI | MRI reconstructive model. | Reconstruct the same MRI with 6x fewer samples. | Can operate on the machine control box with no network crossing server processing. |
If you are comparing against the usual cross modal retrieval stack, this is our answer to it, and we built it to win on quality per watt and performance. It really rips.
FELA Powered Scientific Computing
We had so much success with the FELA methodology that we are proud to demonstrate:
| Product | Description | Use Case / Benefit | Key Advantage |
|---|---|---|---|
| FELA PDE | Solves partial differential equations locally. | Eliminates waiting for thermal simulations on big workstations or cloud compute. | Within 5% accuracy, completely local. |
| FELA PDM | High performance compute predictive maintenance model. | Enables turbines to predict when they will fail. | On device processing, no external dependency. |
| FELA Grid | Solar panel management and optimization model. | Forecasts weather and adjusts panel operations dynamically. | Can operate in the field with no network required. |
| FELA Moderation | Speedy first pass or entirely on client content moderation. | Prevents toxic content from ever being saved to servers. | High speed, client side prevention. |
| FELA Edge | High efficiency time series forecasting model. | Forecasts grid usage for PUDs (Public Utility Districts) or Power Buyers. | HFT for power buyers? Saves money and electricity - efficiency! |
| FELA Genomics | Specialized genomics modeling tool. | Labels promoters without requiring cloud infrastructure or expensive GPUs. | Outperforms HyenaDNA on 4 out of 6 standard tasks. |
| FELA Chem Suite | High speed chemical informatics tool suite. | Scores chemical reactions and determines one step retro precursors in seconds. | Exceptional speed on complex chemical tasks. |
| FELA Tab | FELA version of TabPFN. | Self classifying, regressing data - the AutoML holy grail. | Error bars, more interpretable than opaque box deep solutions. |
Using our high efficiency model, we can enable scientists to move faster and waste less, bringing about a cleaner future with cleaner tech.
(If you have a custom computing concern, chances are, we can find a way to get FELA to make it computable. Contact our enterprise sales for more details.)
FELA LLM
Fourier Encoder with Linear Attention
The world's most efficient sub quadratic language model, distilled into a responsible autocomplete coding model. FELA is built for long context engineering work that stays on your machine, fast even on CPU inference. Open weights with quantized variants, available to self deploy and via API.
The architecture is documented in our research record on Zenodo. We will be updating our progress publicly as it goes through conference submissions.
The best experience with FELA, currently, we feel is the autocomplete model. We aimed to create a chat model that could center humans and put human technical creativity squarely in control.
We feel better about using FELA on our desktops to write code faster. It can read a whole repo into context, and it can suggest useful, well formatted autocomplete suggestions, quickly, privately and most importantly, unintrusively.
(Did you know - we could use funding to attain scale that would make larger variants useful!)
What we stand for
We care about the environment. We build efficient AI on purpose. Our models are designed to run powerfully and cleanly on commodity hardware. The products we are building are right sized. We aim our work at problems worth solving, like models that support climate research and resource efficiency, and understanding our physical world better.
We care about people. Consent comes first. We aim to construct clean datasets with permission and tracked provenance, and we are committed to finding pathways to pay creatives fairly for the work that makes our models possible. No theft baked in.
We will not repeat the same tired patterns. Lowdown Labs is worker managed and operated, moving toward a fully democratic public benefit corporation so our incentives stay aligned with doing good rather than extraction; and so that our workers are equal and empowered every single day. DEI is the core of how we are built, not a footnote.
Licensing
Every model we release uses the Hippocratic License 3.0 + CC BY-NC 4.0. We will never release a model to replace a human or to generate creative work in someone's place. Read that as a promise, not fine print - we do what we do to understand our world, not to cheapen it.
Get involved
- Website and access: gimmelowdown.com
- Status: status.gimmelowdown.com
Think of us as fair trade for AI. Better is possible. This is the way.