Unverified50% confidenceFactExact time
Cloud-based APIs can typically achieve output speeds of tens to hundreds of tokens per second, while consumer-grade GPUs running 27B models often only reach 10–30 tokens per second.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/22/2026
First Seen
Valid until: 12/21/2026
Sources
Related Entities
Related Claims
UnverifiedMainstream GPU inference solutions (such as deployments based on NVIDIA H100) typically achieve only a few dozen to one or two hundred tokens per second on 27B-scale models.80% similarUnverifiedFull-precision video diffusion model inference typically requires a GPU with more than 24 GB of VRAM, whereas 0.4MP with 10-step Turbo acceleration may reduce VRAM requirements to the 12–16 GB level.75% similarUnverifiedOpenAI's GPT-OSS 20B achieved a measured speed of over 21 tokens/second, nearly twice that of the dense 12B model.75% similarUnverifiedAchieving video generation faster than playback most likely involved NVIDIA's latest-generation data center GPUs (such as the H100 or the upcoming B series); on consumer-grade GPUs (such as the RTX 4090), inference speed of the same model may be 5–10 times slower.74% similarUnverified将模型权重从FP16压缩到INT8或INT4的量化技术，可使消费级GPU如RTX 4090在本地部署时达到30-80 TPS的实用速度74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942053API
curl https://kongchang.com/api/v1/knowledge/claims/942053MCP
get_claim(id=942053)