Tencent Hy4-Flash Open-Weights (224B / 12B Active MoE · FP8 & Base) VRAM, 1/24th KV-Cache & API Cost Calculator (2026)
Complete Day-0 hardware VRAM, quantization, and serverless cost calculator for Tencent’s newly open-sourced Hy4-flash-FP8 and Hy4-flash-Base models (224B total / 12B active LatentMoE + 64B Engram, 1M multimodal context, 150+ tok/s, 1/24th KV-cache footprint)—comparing self-hosted 4-bit (~118GB VRAM) vs OpenRouter ($0.035/$0.1405) and Tencent TokenHub ($0.070/$0.281, $0.003 cached).
Live 2026 FinOps Benchmark Matrix: Open-Weights Multimodal & Agentic FinOps
| Architecture / Model Tier | 2026 Unit Economics & Pricing | Production FinOps Impact |
|---|---|---|
| 224B Total / 12B Active LatentMoE (+ 64B Engram) | FP8 Checkpoint: ~228GB VRAM · 4-bit AWQ/Q4: ~118GB VRAM · 1M Multimodal Context | 4-bit fits cleanly inside 2x H100 80GB, 2x DGX Spark (256GB), or a single NVIDIA B300 node |
| 1/24th KV-Cache Compression & 150+ tok/s Speed | Reduces KV-cache memory to 1/24th of Hy3 and 1/6th of Hy4 via DFly & megakernels | Prevents VRAM OOM during 20-turn 200K–1M token coding and document-extraction loops |
| Serverless API Pricing Floor ($0.003/1M Cache Read!) | $0.035/$0.1405 on OpenRouter (50% promo) · $0.070/$0.281 official list ($0.003 cached) | ~18x lower end-to-end task cost than Claude Haiku 5.5 ($0.80/$4.00) on daily productivity suites |
| Self-Hosted GPU Break-Even Threshold | Serverless wins below ~45M tokens/day; dedicated 2x H100 / 1x B300 wins above ~60M/day | Filter all 224B–320B sparse MoE models in 1 click using TokenMargin’s <=256GB VRAM button |
Related FinOps Playbooks & Model Cost Calculators
Drop Your Usage CSV, Paste Your System Prompt, or Rescue Your Monthly LLM Bill
100% browser-local FinOps auditing: drop your OpenRouter / OpenAI / Anthropic / Cursor usage CSV, count 20-turn agent loop bloat from your system prompt, or model instant savings across 61 models.
Export your 30-day activity CSV from OpenRouter, Anthropic Console, OpenAI Usage, or Cursor/Cline and drop it below (or paste raw CSV rows). Everything runs in your browser memory — no API keys or logs ever leave your device.