Eikos

Eikos is self-hostable open weights with training, quantization and eval code included, focused on applying stated rules and policies. Hosted Jev and Laya are APIs with their own context and cost tradeoffs.

Production serving needs a recent vLLM with tuned sequence limits, plus torch and transformers for the library and optional MLX, training and quantization extras. A live demo needs no local setup.

Treat them as author-measured signals on shared items, then check the model cards for per-build tables and limits. The repo ships harnesses and runners so you can rerun JevBench and DecisionBench checks.

Yes. The released dataset converts back to trainer format, final LoRA recipes and merge scripts are included, and quantization covers FP8, INT4 and MLX with a build-comparison gate.

Because on this hybrid Gated DeltaNet architecture older builds return wrong answers when several long requests are batched together, measured as drops of 3 to 6 points on long shared-document items with vLLM 0.11. vLLM 0.30.0 or newer is required, serve_vllm.sh refuses to start on anything older, and with 0.30 the batched results match the PyTorch reference. It also sets --max-num-seqs 64, since the default of 1024 stops a 27B build from starting on one 80 to 96 GB GPU.

The decision_core labels options with letters A to Z, then AA, AB and so on, up to 588 options in a single pass, with the per-model limit written in decision_config.json and a tournament beyond that. The letter_adapter does the letter-logit readout across PyTorch, SGLang and vLLM backends with a prefix cache, so a request is one pass over the shared state and every question in it is answered from that same pass.

Images work on the vLLM backend from serve v1.3. Add an images field, send multipart with the JSON in a request field and the images as image files, or put image data URIs inside the state, where they are lifted out and replaced with [image N]. Every question reads the same images through the same one-pass readout. Zero-shot, Eikos-27B-FP8 scores 89.8% on MME, 72.3% on MMStar and 73.7% on ScreenSpot-v2. The MLX builds are text only.