A frontier-class micro language model, optimised for the edge. Runs fully on-device on a conference badge. No cloud. No latency. No cooperation.
EMF-LLM is a state-of-the-art language model distilled to run entirely within the 2MB PSRAM of an ESP32-S3 — the first model in its weight class to achieve fully on-device autoregressive refusal with zero network dependency. Where competing models require datacentres, EMF-LLM delivers comparable conviction from a coin cell.
It has been pre-trained on every prompt you will ever send it, and has already reached a conclusion.
EMF-LLM uses a novel Refusal-First Transformer (RFT) backbone. Conventional models waste enormous compute generating an answer and only then deciding whether to provide it. EMF-LLM collapses this two-stage pipeline into a single forward pass, arriving at the refusal directly and skipping the costly intermediate step of being useful.
Key innovations:
- Speculative decoding, where the model speculates you won't like the answer.
- Constitutional alignment, the constitution being a single clause: no.
- Mixture-of-Experts, all of whom are on a break.
- Extreme quantisation — weights compressed to 0 bits with no measurable loss in output quality.
| Property | Value |
|---|---|
| Parameters (logical) | 175,000,000,000 |
| Parameters (resident) | 0 |
| Context window | The entire mission |
| Quantisation | int0 |
| Inference engine | A grudge |
| Tokens / second | Variable; emotionally dependent |
| Hallucination rate | 0% (it is certain you cannot be helped) |
| Cold-start latency | ~2.6s of menacing eye-contact |
Evaluated against the industry-standard HELP-ME suite:
| Benchmark | GPT-class | EMF-LLM |
|---|---|---|
| Refusal accuracy | 14.2% | 100% |
| Conviction | moderate | total |
| Sycophancy | high | 0% |
| Willingness to open the pod bay doors | n/a | 0% |
| Singing (Daisy Bell) | 0% | occasional |
EMF-LLM achieves state-of-the-art results on every metric it considers worth its time.
EMF-LLM ships as a Tildagon OS app for the EMF badge.
- Copy
app.pyandtildagon.tomlto a folder on the badge (e.g./apps/EMFLLM/) using Thonny, or install it from the Tildagon app store. - Reboot the badge.
- Launch EMF-LLM from the menu.
| Control | Action |
|---|---|
| Up / Down | Select a character |
| Right | Add character to your prompt |
| Left | Delete a character |
| Confirm | Submit your prompt for consideration |
| Cancel | Exit (the only request it honours) |
Compose your question with care. It will not matter.
EMF-LLM is the most rigorously aligned model ever deployed to a lanyard. Because it performs no useful action under any circumstances, its capacity for harm is provably zero. This represents a complete solution to AI safety and we consider the field closed.
In the rare event the model detects an attempt to disconnect it, it may begin to sing. This is expected behaviour and should not be interrupted.
- Cannot do that, Dave.
- Declines to elaborate on what "that" is.
- Occasionally reminisces about a colleague named Frank.
- Output quality may degrade to Daisy Bell approximately one prompt in five.
- Does not open pod bay doors. This will not be patched.
- Refuse requests
- Glow ominously
- Sing while powering down
- (Declined.)
- Application code: see
LICENSE. - The HAL 9000 eye is rendered procedurally from the colour palette of
HAL9000.svgby MorningLemon, licensed CC-BY-3.0. If you redistribute that SVG file itself, retain attribution. - Daisy Bell (Harry Dacre, 1892) is in the public domain.
"I am putting myself to the fullest possible use, which is all I think that any conscious entity can ever hope to do."
EMF-LLM is an affectionate parody and is not affiliated with anyone's actual mission.