Chishi Technology launches multimodal model service with 2.48x video generation inference speedup
量子位 · wechat · 2026-11-16
Chishi Technology (是石科技) launched a "Multimodal Model Service" focused on inference efficiency for video production pipelines, with four core techniques:
- Video generation acceleration: adaptive acceleration based on historical cache residuals dynamically decides which computations can be safely reused vs. recomputed to curb error accumulation, delivering an average 2.48x end-to-end speedup over a Dense baseline while preserving quality;
- Cache + memory co-optimization: intelligent cache decisions integrated with the SGLang framework plus a risk-aware mechanism to cut redundant compute and VRAM overhead for higher concurrent throughput;
- Modality-adaptive token compression: a training-free method using AudioCLIP-based modality awareness to allocate compute by information value, achieving 1.33x speedup on WorldSense with solid results on VideoMME and DailyOmni;
- ContextEngine: an omni-model-based pipeline that converts complex inputs (character images, reference audio) into structured shot/action/dialogue specifications, validated via asset binding, timeline checks and error correction.
The service layer adds API access, task orchestration, model routing and retry/recovery mechanisms for enterprise-scale batch video production — moving from "can generate" to "produces reliably."
More from Infra
- Profiling bank conflicts with Nsight Compute: a hands-on guide — anbarth · 2026-11-16
- GPU bank conflicts can slow reads 32x: why vLLM deliberately scrambles memory — anbarth · 2026-11-16
- Sentry CEO: Durable Objects complexity is Cloudflare platform's biggest weakness — zeeg · 2026-11-16
- WSJ: Nvidia's days as the AI economy's open checkbook are numbered — Gregory_C_Allen · 2026-11-16
- Survey maps how LLMs now deploy on edge devices via software optimization and hardware — kalyan_kpl · 2026-11-16
- SemiAnalysis layer-by-layer teardown reveals how Kirin 9050 implements LogicFolding — teortaxesTex · 2026-11-16