Unverified50% confidenceReleaseExact time
Comfy-Org provides a LoRA extracted from Turbo (approximately 0.91 GB) that can be applied to the base DiT without downloading the full 14 GB checkpoint.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/11/2026
First Seen
Valid until: 1/9/2027
Sources
Related Entities
Related Claims
UnverifiedThe 24 GB GPU streams the 34 GB DiT by loading weights through ComfyUI's dynamic VRAM path, actually occupying 23 GB of VRAM, at the cost of consuming 68 GB of host memory64% similarUnverifiedUsers with an NVIDIA RTX 4090 (24 GB VRAM) can run 70B-parameter models locally for personal experimentation and development via 4-bit quantized versions in GGUF format.63% similarUnverifiedThe Ascend 950PR is equipped with 128 GB of VRAM.62% similarUnverifiedThe full weights of LLaMA 2 (70B parameter version) amount to approximately 130 GB.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/996810API
curl https://kongchang.com/api/v1/knowledge/claims/996810MCP
get_claim(id=996810)