2026-10-11 17:13 UTC

Independent testing will determine whether Ullis can train and serve ternary MoE models on local hardware with useful correctness, performance, and memory efficiency.

state: expiredheat: lowuncertainty: highknownscott: mediumlocal-inference open-models inference-economicsVladislav Kalinkin

What is this?

Ullis is described in the supplied evidence titles as an open-source Rust engine for locally training and serving ternary mixture-of-experts/KAN models, associated with Vladislav Kalinkin. The search snippets establish only the broader context: MoE architectures can reduce FLOP requirements but introduce routing, expert-collapse, memory, and hardware-optimization challenges, while quantization and offloading can improve local viability. None of the supplied snippets independently tests or even directly discusses Ullis, so its correctness, speed, memory efficiency, and practical local-hardware performance remain unverified here despite the web answer’s unsupported claim of confirmation.

Why it matters to Scott

This is another unverified implementation in territory already tracked by the radar through `radar:bonsai-extreme-quantization`, `radar:maple-ternary-iphone-inference`, and related local-inference cases. It matters beyond topical fit because Scott could benchmark Ullis directly on `dev:project.gamepc` under his `dev:concept.hardware-aware-local-inference` and evaluation discipline, but no supplied evidence yet shows results that would change his position or stack.
dev:concept.hardware-aware-local-inferencedev:project.gamepcip:concept.evaluation-driven-developmentradar:bonsai-extreme-quantizationradar:maple-ternary-iphone-inferenceradar:concept.local-inferenceradar:concept.extreme-quantizationradar:concept.moe-inference
queries asked of Scott's wikis
  • local model training and inference economics
  • ternary weights and extreme quantization
  • sparse MoE routing on commodity hardware
  • Rust runtimes for local AI systems
  • open-model sovereignty through local training
  • benchmarks for local inference correctness and memory

Measured heat

no measured readings yet β€” the hourly heat pass fills this in

How the heat travelled

no chain yet β€” the hourly chain pass fills this in

Evidence (2) β€” ⭐ canonical anchor

sourceobjectauthorscorecomments
🟧 hnShow HN: Ullis – Local Ternary Moe-Kan Training and Inference Engine in Rustvlad_kalinkin10
🟧 echo.github ⭐An open-source Rust engine for local ternary MoE-KAN training and inference.Vladislav Kalinkinβ€”β€”

Interpretation history

Decision trace