Show HN: Hops Run: beat Semif (open Jev) and Kumo (Nvidia) in a procedural game

Category: ai-ml

Tags: semantic-decisions, llm-inference, logits, calibration, benchmarking

Score: 7.3/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 6)

SemIf is an open-source reimplementation of the semantic decision interface pattern from TypeSafe's closed Jev service, using open models (Qwen, MiniCPM) to read typed option probabilities directly from logits without generating text. It offers multiple backends (CUDA, MLX, llama.cpp) and demonstrates up to 5x speedup over autoregressive JSON generation, with committed benchmarks and calibration. Interesting as a practical, auditable baseline for fast, structured LLM decisions in agent pipelines.

Target audience: AI/ML engineers, backend developers building agent systems, researchers

Repository: https://game.hopsworks.ai/ · Python · MIT · 4876 stars

View on Hacker News