DeltaTensors Stores Fine-Tunes as Weight Deltas: 953MB → 294MB With Minimal Quality Loss
cupheadgamer · reddit · 2026-09-22
Open-source DeltaTensors manages fine-tuned checkpoints like Git, storing only the weight delta from the base model instead of full copies. On a Qwen2.5-0.5B fine-tune, a 953MB model compressed to 294MB with perplexity moving only 19.11 → 19.22. It works post-training (no LoRA or workflow change needed), supports chained deltas for version history, Hugging Face Trainer checkpoints, and streaming compression so you never load two full models. Repo: AaravGaurdev/deltatensors.
More from Infra
- Gas-for-AI-datacenters forum co-located with CoRL shows energy demand spillover — carlosdponx · 2026-11-12
- a16z podcast: Mike Volpi on Cisco's 70+ acquisition playbook and the AI buildout — a16z · 2026-11-12
- AMD to pass Intel in server revenue by mid-2027 as share gains 6 points a year — BenBajarin · 2026-11-12
- Study: Reflective Data Center Roofs Actually Cool Daytime Ground by 1.8-2.2°C — SumitGup · 2026-11-12
- Tsinghua's BaGuaLu-2 tames MoE expert-load spikes at runtime, accepted to SC26 — jiqizhixin · 2026-11-12
- antirez backs low-bit DeepSeek 4.1 Flash over Qwen 27B dense on Apple M5 Macs — antirez · 2026-11-12