VIDRAFT This Week: Typed Decisions #1, S1MB Japanese #1, and NPU Pretraining in the Press
aillmbenchmarksnpu
TL;DR
Three results from VIDRAFT this week, each one checkable on a public page:
- Typed Decisions: Darwin-27B-ZTC-v5 is #1 on the Hugging Face typed-decisions leaderboard at 0.81 accuracy. The earlier Darwin-27B-ZTC holds #2 at 0.74.
- S1MB Japanese edition: #1 at 84.57, measured by the benchmark operator.
- VKORE: pretraining a language model on inference-only NPUs, covered by JoongAng, ChosunBiz, ZDNet Korea and AI Times, and internationally by TechGig and Mukta AI.
Typed Decisions
Typed Decisions scores how well a model makes structured choices: pick the right option, return a typed value, decide whether an action should run. It is the format agents use when a decision has to be machine-readable rather than conversational.
The current leaderboard top four:
| Rank | Model | Accuracy |
|---|---|---|
| 1 | FINAL-Bench/Darwin-27B-ZTC-v5 | 0.81 |
| 2 | FINAL-Bench/Darwin-27B-ZTC | 0.74 |
| 3 | empiriolabsai/aplomb-1 | 0.74 |
| 4 | openjev/openjev | 0.73 |
Source: LocalLLaMA/typed-decisions.
S1MB, Japanese edition
The System One Mosaic Benchmark aggregates many decision-style tasks into one ranking. On the Japanese edition, our model ranks first at 84.57. The score was measured by the benchmark operator, not self-reported.
VKORE in the press
VKORE is our acceleration layer that lets inference-only NPUs run training workloads. This week national outlets in Korea reported on it, alongside TechGig and Mukta AI internationally. The point the coverage picked up is practical: training capacity that does not depend on scarce GPUs.
What connects them
Decision-grade models on one side, and the infrastructure to train them without GPU dependence on the other. We will keep publishing results as they are measured.