Unverified60% confidenceFactExact time
BitNet b1.58 uses ternary quantization, where each weight takes one of three values—-1, 0, or +1—amounting to approximately 1.58 bits (log₂3).
2
Sources
60%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBitNet 1.58b quantization uses ternary weights {-1, 0, +1} in PCEI's 24-layer network, reducing parameter memory by about 4 times compared with FP16 precision, and at inference time integer addition can replace floating-point multiplication.76% similarUnverifiedThe follow-up paper "The Era of 1-bit LLMs (BitNet b1.58)" found that pure 1-bit requires very large scale to match full-precision quality, so it introduced a zero state, with weights becoming -1, 0 or +1.73% similarUnverified三值化权重每个权重仅携带约1.58 bits信息量（log₂3≈1.585），相比FP32的32 bits信息密度下降约20倍72% similarUnverifiedFor a model with exactly 8 billion parameters, the idealized weight payload is about 14.90 GiB at 16 bits, about 7.45 GiB at 8 bits, and about 3.73 GiB at 4 bits.65% similarUnverifiedUnder 2-bit quantization, each weight has only about 4 discrete values.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/795618API
curl https://kongchang.com/api/v1/knowledge/claims/795618MCP
get_claim(id=795618)