Most "AI-native" chains call a commercial model API and put a logo on it. The Empoorio ecosystem trains its own. This post is about EmpoorioLM — what the models are, where they run, how they are evaluated — with the capability level stated plainly, because small models are small.
The family
| Model | Parameters | Tokenizer | Purpose |
|---|---|---|---|
empoorio_lm_small_v1 | ~25M | small_bpe_v1 | General text; the base for the LoRA adapters the network trains |
empoorio_lm_code_40m_v2 | 40M | tokenizer_v3_code | Code |
empoorio_lm_60m_v1 / 64M | 64M | tokenizer_v3_code | Larger code and text base; the current mobile-training base |
All are trained from scratch by the ecosystem; the architecture, training guide and scaling roadmap (toward 3B+ parameters) are documented in the Ailoos repository. The decision to bet on native models rather than a fine-tuned third-party base was made explicitly in August 2026.
Where they run
api.ailoos.com serves them. Two paths: the coordinator loads EmpoorioLM checkpoints directly with PyTorch and exposes an OpenAI-compatible endpoint; a separate llama-server on the host serves a quantized general model (AndromedaLM) that the browser Aephoron and the Warehouse app use as an assistant. A health check on the API returns healthy; it was measured on 8 September 2026.
The tokenizer is part of the model
The most instructive production bug of 2026: the inference server used one global tokenizer for all three models. Measured id by id, small_bpe_v1 and tokenizer_v3_code agree on 275 of 8,000 ids (3.4 %). Decoding with the wrong one raises no error — it produces fluent, plausible, wrong text, and the symptom reads as "the model is bad". The fix pairs each model with its own tokenizer, derives stop tokens from it (the old hardcoded stop id meant '<|' in one vocabulary and 'exp' in the other, truncating code output at the first "exp"), and refuses to load a model whose tokenizer is missing rather than falling back to the global one. Any constant derived from a tokenizer travels with it.
Evaluation
EmpoorioLM is evaluated on held-out loss and on task probes appropriate to its size. There is no benchmark claiming it builds complex on-chain transactions from natural language; the honest statement is that a 25–64M model completes short code and text, follows simple instructions unevenly, and is the right size to train on a phone — which is the point. The Warehouse app's inventory notes its assistant's output quality as "unreliable at 1B, with a validation guard"; the smaller EmpoorioLM models are used where the loop matters more than the eloquence.
The base matters
When the production mobile-training artefact was audited, it turned out to contain random weights — every exporter in the repository built the model and exported it without loading a checkpoint, and mobile nodes had been fine-tuning LoRA adapters on noise. The fixed exporter verifies that the real weights appear inside the artefact before writing it. On the random 40M base, training loss hovered around 9.0; on the real 64M base it starts near 6.9 and falls to 6.4 in 30 steps on a phone. That gap is what a real model looks like.
What EmpoorioLM is for
A model the network can improve collectively, on consumer hardware, with provenance for every dataset that trained it (a manifest records which of the 421 catalogued datasets went into each model) and rewards paid in DMS per era for verified work. Not the largest model — the one that closes the loop.
Based on EMPOORIO_LM_ARCHITECTURE.md, MODEL_SCALING_ROADMAP.md, the inference-server fixes of 2026-09-08 and the mobile-training verification.