AI Briefing#compression
The latest AI and developer news about #compression, with the original source and a short summary.
The latest AI and developer news about #compression, with the original source and a short summary.
The framework scales group size quadratically with spatial compression to enable 32x compression while maintaining competitive performance on five text-to-image benchmarks.
1-bit TurboQuant compression reduces vector size from 856 KB to 40 KB per page while retaining 96.5% of float32 retrieval scores.
Recent pipeline changes reduced backbone traffic by ~70 GB/s, saving significant CPU and memory resources.
Based on Qwen3.5-9B, LensVLM maintains original-level accuracy at a 4.3x compression rate and demonstrates performance up to 10.1x.
LLM research agents generalize without benchmark overfitting because their successful strategies possess high compressibility.
Cloudflare has unveiled a prototype that applies Zstandard compression to triple cache storage capacity and reduce bandwidth usage.
This introduces Speculative KV coding, a technique that uses a predictor model to losslessly compress an LLM's KV cache by up to 4x.
Shard has been released, compressing the KV cache memory of Llama-3.1-8B by about 10x without performance degradation.
Compressing Python source with n-grams and arithmetic coding achieved a 33% improvement over zlib.
Splitting core/scratch compresses reasoning KV cache smaller and more accurately.
TriAttention compresses the KV cache by 10.7x and improves throughput by 2.5x through trigonometric-based compression without loss of accuracy.
NVIDIA has released KVPress, a KV cache compression toolkit for handling long context in LLMs.
AI21's Summarize API produced more accurate and condensed summaries than OpenAI models.