dsh-models-radar

Switching between models and reasoning tiers in DeepSeek Harness often leaves you juggling tabs on external benchmark sites, guessing at capability differences until the invoice arrives at month-end. dsh-models-radar pulls the public benchmark data from deng.codexradar.com directly into the Settings page, placing capability scores, trend curves, cost-efficiency maps, and community ratings on a single screen.

The core experience works in three layers. The capability overview groups results by base model on a fixed 0–110 IQ scale, with each row expandable to reveal the full reasoning-effort ladder. Dual-window trend charts (24 h and 7 d) are independently y-scaled and include net change, mean, and extreme values; the curve is colored by capability band. The cost-by-IQ card offers composite, time, and price tabs on a log-cost axis, where the upper-left corner marks the most efficient tiers and same-base tiers are joined by ladder lines. Community ratings add a rolling 7-day and 24-hour 0–10 experience score for a subjective reference layer.

A live capsule beside the model selector in the composer updates the instant you switch models, and clicking it opens a cross-base comparison popover without leaving the editing flow. The plugin requires no credentials, tolerates offline use with a local snapshot fallback, and never sends session content upstream. It suits developers who weigh efficiency trade-offs across Codex, Claude Code, Grok, and other harnesses, as well as architects who want to demonstrate how far a single base model stretches across reasoning tiers.