Eikos weights, code and data are out: open 4B and 27B decision models from Brazil
Hours after announcing it in Portuguese on X (@0xCVYH, 23 Sep 2026, 12:09 BRT), Brazilian developer Caio Vicentino put the whole Eikos release online the same evening. Eikos-4B and Eikos-27B landed on Hugging Face between 20:48 and 20:59 BRT with FP8 and INT4 builds, MLX builds for Apple Silicon followed on 24 Sep, and the GitHub repo and the eikos-decisions dataset went up alongside. We only caught it on 26 Sep; the catalog page had been saying “release pending” in the meantime.
What shipped. Two LoRA fine-tunes (rank 64, one epoch) merged into full checkpoints: the 4B on Qwen3.5-4B and the 27B on Qwen3.8-27B. They answer noul, choice and score questions in one pass by reading option-letter logits (SemIf prompt format), with a probability per option. serve.py exposes a TypeSafe-compatible POST /v1/systemone plus agent sessions; vLLM 0.30 or newer is required. The repo includes the data pipeline, training recipes, quantization gate and evaluation harness. Licensing: MIT for the fine-tuning deltas and code, Apache-2.0 for the Qwen bases, CC BY 4.0 for the dataset. The focus is rules in finance, trading and trade finance.
What is claimed (UnverifiedClaim). 82.9% on the hard tier of the public JevBench items for the 27B, against 73.0% for Jev measured by the author through its API; about 2.4% error on decisions taken at ≥90% confidence; ECE around 0.04; 88.3% with a decision buried in 64k tokens. These come from the author’s own harness. The official JevBench leaderboard entry is still pending, the rules suites share a generator family with the training data, and on the author’s general battery Jev still comes out ahead (84.1 vs 82.5).
Unlike most launches in this category, the full recipe is public, so an outside rerun is possible. None has been published yet.
Catalog entry (updated): /models/eikos. Related: SemIf (same readout on frozen models), Jev.