Skip to content

NeuralDash v0.5.0

Choose a tag to compare

@Legesus Legesus released this 29 Oct 06:26
· 15 commits to main since this release

Benchmarks, a sidebar chart, and redesigned hover cards:

  • Live benchmark scores — rows are scored with the Epoch Capabilities Index, read at most daily from Epoch AI's benchmark download (https://epoch.ai/data/benchmark_data.zip, eci_scores.csv, CC BY 4.0). Board models are matched by normalized slug, display name, then family; ambiguous or unknown models stay no benchmark, and fast/speed variants inherit their family score labelled (family). Value ranking (score ÷ mWh) now works out of the box. Bundled and override scores still win per slug.
  • 48h trend in the sidebar — the chart moved from an editor tab to a 48h Trend pane under the board. Clicking a model row (or the chart action / tooltip link) shows it there, and it re-renders on every board refresh. The chart is now an inline SVG with 7-day typical and 48h-average reference lines, hi/lo markers, per-point hover values and a low/median/high strip.
  • Redesigned model hover card — headline energy with the board's trend badge, a compact energy table, a text sparkline linking to the chart, the top prompt-size bands ranked by traffic share, and an Updated … · source footer.
  • Flex card: Normal → Flex comparison — per-token price table (input, output, cache read) and a charged-energy table (energy per request and cost per 1k requests), both computed from the flex_discount_multiplier in https://api.neuralwatt.com/v1/models, which per Neuralwatt's docs applies to token billing, pay-as-you-go energy and subscription allowance drawdown. Provenance collapsed to a single Sources: line; full detail stays in NeuralDash: Debug Flex.
  • Board markup fixes — the energy board now groups models under family headers: rows are named from the header (Kimi K3 Fast, DeepSeek V4 Flash · Speed lane) and keyed by their real model id; the context band is read from any sub-line segment; 48h trend direction reads the new nw-spark-cheaper / nw-spark-pricier classes.
  • Cost per 1k requests shows 2 decimals; flex token prices keep up to 4 so sub-cent prices don't round to zero.
  • CI: gitleaks secret scan on push and pull requests.