AnalysisAI ModelsSeptember 16, 2026

Intel's BITCOS format compresses ternary LLM to 1.485 bits

Intel researchers compressed a 1.58-bit ternary model checkpoint to 1.485 bits per weight using a new format called BITCOS, without altering any model weights. The gain comes from changing how weights are stored, and decoding also improved.

2 sources

More stories today

Open the live feed