Horizon
A verification-first layer for local LLMs that adds hard test execution, symbolic math checking, and an agentic repair loop on top of a small open-weight model, aiming to match a much larger teacher model's accuracy on checkable tasks.
Horizon is a research project arguing that the model itself is not the differentiator for local coding assistance; the verification layer around it is. It wraps an open-weight language model running on a 16 gigabyte consumer machine with a metacognitive router, an execution harness that runs sandboxed tests and symbolic math checks, best-of-N answer selection, and an agentic repair loop for code, so that every answer ships with evidence that it was actually tested rather than only generated.
On tasks where correctness can be mechanically checked, such as code with tests or verifiable math problems, the project reports a 7-billion-parameter model reaching parity with, and on a newer base model exceeding, its own 671-billion-parameter teacher model on a code-generation benchmark, while running at about 13 tokens per second on ordinary desktop hardware. The layer is training-free at inference time and is designed to reapply to any open base model without modification, so results are also replicated across additional model families.