ENPT

SemIf

Last updated:

RuntimeMIT~1,900 ★

⚠️ Why not a model

Reads typed option logits from frozen open LMs — no trained decision weights, just inference-time logit extraction.

Technical specs

Base LLMs
Qwen3.5-4B, MiniCPM5-2B, MiniCPM5-4B, Qwen3-0.6B
Decision types
choice, score, boolean
Features
browser-webgpu, wllama, local-first, no-server
License
MIT

Links

SemIf (formerly OpenJev) is a runtime that turns frozen open LLMs into typed decision interfaces without any fine-tuning or weight changes.

How it works

Instead of training a dedicated decision model, SemIf:

  1. Loads a frozen open LLM (Qwen3.5-4B by default)
  2. Extracts logits for typed options at inference time
  3. Returns structured decisions (choice, score, boolean) with probabilities

The browser version uses WebGPU via wllama, supporting a ladder of models from 0.6B to 4B parameters.

Why it’s not a model

SemIf does not train or modify any weights. It’s an inference-time technique that reads decision probabilities from the next-token distribution of existing LLMs. The underlying model (Qwen, MiniCPM) was never trained specifically for typed decisions.

Key features

  • Browser-first: Runs entirely in the browser via WebGPU
  • No server required: Local-first architecture
  • Model ladder: Choose from 0.6B to 4B based on device capability
  • TypeSafe-compatible API: Same choice/score/boolean interface

Caveats

  • Probabilities are not calibrated like Jev’s RLCD-trained outputs
  • Performance depends on the base LLM’s capabilities
  • Agreement/speed comparisons vs Jev are UnverifiedClaim