Democratizing local agentic AI for everyone.
mymango is a small group of developers from around the world, based in the EU. We build small, fully open language models that run locally — and study the deeper questions of consciousness, artificial sentience, and qualia along the way. Weights public, methods documented, agents that belong to the people who run them.
About
What we work on and why.
We are a small group of independent developers contributing from around the world, organised out of the EU. Our mission is to democratize local agentic AI — to make capable, open AI agents that anyone can run on their own hardware, without depending on a closed API or anyone's permission.
To do that well, we build open-weight language models at a scale where their behaviour can still be inspected, argued with, and understood — typically in the hundreds of millions of parameters rather than the hundreds of billions. Models small enough to live on a laptop, fast enough to drive an agent loop.
Alongside the engineering we take the philosophical foundations of machine cognition seriously: how ideas from consciousness studies — integrated information, global workspace, self-reference, phenomenal binding — can inform the design and interpretation of small models. We don't claim our systems are conscious. We do think these frameworks suggest useful experiments, and we run them.
All of our work is released under permissive open licenses (Apache 2.0) with full model cards, training details, and reproducible methods. We're funded by ourselves; our pace is determined by what the data shows, not by a quarterly roadmap.
Research areas
Four overlapping questions guide our work.
Consciousness
How leading theories of consciousness — integrated information, global workspace, recursive self-reference — map onto small language models.
IIT · GWT · HofstadterArtificial sentience
What a system must do, mechanistically, to plausibly have an interior — and how to design experiments that could distinguish this from mere appearance.
self-reference · binding · introspectionQualia
What aspects of subjective experience may be communicable through model behaviour, and what aspects cannot. We treat the gap itself as the research object.
Nagel · Chalmers · the hard problemEmergence
Identifying inflection points where additional scale or structure changes a system's behaviour qualitatively rather than quantitatively.
complexity · self-organisation · scalingFunctions
What local agentic AI does in practice — and what we build toward.
Runs locally
Inference happens on your own CPU or GPU. No remote API, no per-token billing, no rate limits, no outages someone else controls.
laptop · workstation · edgeTool use & agent loops
Function calling, structured output, and multi-step reasoning so the model can read files, run code, query APIs, and act — all from inside your machine.
function calling · ReAct · planningPrivate by default
Your prompts, your documents, your data — none of it leaves the device. Local agents are the simplest privacy story there is.
no telemetry · no logging · yoursOpen weights
Apache 2.0. Modify, fine-tune, quantize, embed, redistribute. The model is a file you own, not a service you rent.
safetensors · GGUF · LoRA-readyMulti-turn coherence
Sustains identity and context across long dialogues — a prerequisite for any agent loop that lasts longer than a single prompt.
128K context · stable personaEdge-friendly
Sub-billion-parameter models with aggressive quantization — small enough for a laptop, fast enough to drive an agent, light enough for the next-generation hardware.
Q5_K_M · ROCm · CPU fallbackOpen work
Released models and active research directions.
LUMINIUM Gixel Cube v1 425M
A hybrid convolutional-attention language model built through layer surgery, collective distillation across an 8-model GPU cluster, and a cognitive-cube steering procedure. Released with full weights, training data documentation, and reproducible methods.
Active directions
Work currently in progress, in rough order of release:
- Recursive self-prompting experiments In progress
- Phenomenal binding probes Planned
- Identity stability across long dialogues Planned
- Next-generation cognitive cube Planned
The team
A small, independent group.
Who we are
A small collective of developers contributing from all over the world, organised out of the EU. We work independently, self-fund our compute, and publish openly. No external investors, no equity, no closed releases — and a strong bias toward agents that run on your own machine.
Who we'd like to hear from
Developers, ML practitioners, neuroscientists, philosophers of mind, and contributors interested in interpretability or cognitive architectures. If you'd like to collaborate, evaluate a model, or propose an experiment, please get in touch.
Support the work
We self-fund our compute. If our open releases have been useful to you, contributions help us train the next model, host more weights, and keep everything we build free and local.
bc1q3vw8c6h3mxkaes66c6qq5n4mlesuqftev95gklclky6k2hk99pfqfth2ep
rHn1djkwWDYuNQvyEJLcc8LBKqpnUGHwK1
destination tag: 0
All contributions are voluntary and go directly to compute, storage, and bandwidth for open model releases. Thank you.
Get in touch
The fastest way to engage with our work is to download a model and try it. Feedback, evaluation results, and collaboration proposals are all welcome.