中

Changelog

Every data and feature change, newest first

2026-09-28

  • Workflow library launched: ten curated H3 recipes (official templates / low-VRAM GGUF / FastH3 VSA / PDD / SLA long clips / Parasyte / batch caching), one page each (/en/workflows/) — required model combos, minimum workable GPU, 864×480 reference times and sources;
  • LoRA library launched: nine step-distilled options, one page each (/en/lora/) — arena ranks, 864×480 reference times per GPU, workable-GPU lists and official downloads;
  • Data refresh (2026-09-28): arena leaderboard refreshed (votes +2%, reshuffle — larry v4 now t2v #1 and i2v #2); ten new models added to the library (Unsloth GGUF, latent upscaler, character-swap / 360° orbit / novel-view LoRAs, Cobijada hybrid pruned, Veda 8NFE, ORB360) and all download stats updated
  • Local cost calculator: power / SSD wear / optional depreciation estimator below the calculator, linked to the selected tier;
  • Two new articles: Prompt Guide and Local vs Cloud Cost Math; new Ad & Affiliate Disclosure page; “Reading the numbers” notes added to all 49 GPU pages
  • Cloud-compute sponsorship slots (RunningHub): official-logo promo banners at the top of the homepage, model library, GPU pages, workflow pages and cloud-cost articles — two slots side by side: the China site (runninghub.cn) and the Global site (runninghub.ai), both shown on Chinese and English pages (China site first on Chinese pages, Global site first on English pages; labelled “Promoted”, rel="sponsored"; sign up via our link for 1,000 free RH Coins) — rules in the disclosure
  • Mobile access reliability fix: removed the inactive AdSense loader (that domain can hang for a long time on some networks and stall first paint); WeChat / Xiaohongshu in-app browsers no longer auto-switch language (their language reporting is unreliable and sent Chinese users to the English site)
  • White-screen fix: some returning visitors got a blank calculator — locally saved numeric settings (electricity price, wattage, …) were mis-parsed into objects and broke rendering. Values are now type-checked and auto-repaired; an error boundary now shows the cause with a one-click “reset site data” button instead of a blank page

2026-09-20

  • GPU library 43 → 49: six modded cards added (RTX 2080 Ti 22G / 3070 16G / 3080 20G / 4080 32G / 4080 Super 32G / 4090 48G) — sources in the data repo
  • H3 Acceleration Arena leaderboard integrated: the calculator now shows i2v / t2v ranks (with votes and source) for each acceleration option; new article “Which LoRA Acceleration Is Right for You?”
  • Open data: full dataset and validation tooling published at GitHub (whichh3-data)
  • Calculator redesign: single-entry flow (pick your GPU, get the answer), fine-tuning moved into “Full config”; added a shareable config link and a weekly update reminder

2026-09-19

  • Internal links switched to clean URLs (removing “Page with redirect” states in search engines)
  • Data date stamps are now generated dynamically
  • Acceleration options updated: Sol-Attn removed, SLA / VSA sparse attention added (with node requirements and conflict hints)

2026-09-18

  • Full refresh of acceleration, model-combo and conflict data; anchor regression mean error 13.9%
  • GPU guide (one page per card) and file reference (one page per file) launched, bilingual
  • 8 tutorial articles published (hardware / quantization / acceleration / troubleshooting / cloud vs local)

Early September 2026

  • WhichH3 launched: H3 hardware fit calculator + model library + data notes

Want updates in your inbox? Subscribe to the weekly reminder at the bottom of the calculator, or Watch the data repo.

H3 update alerts · at most weekly New LoRAs, new GPUs, leaderboard moves — one email