Unverified50% confidenceFactExact time
1-bit量化研究可追溯至微软研究院2024年发布的BitNet论文，提出用三值权重（-1, 0, 1）替代全精度浮点数，BitNet b1.58进一步证明1.58-bit在从头训练场景下的可行性
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified1-bit量化（如BitNet b1.58）将每个参数用-1、0、+1三值表示，目前主要停留在研究阶段78% similarUnverifiedApart from bitnet-b1.58-2B-4T, there is currently no precedent for a properly trained ternary model trained from scratch.71% similarUnverified研究表明当模型参数规模达到3B以上时，1.58-bit模型在多项基准测试中表现可接近甚至匹配同等规模的全精度模型66% similarUnverified4-bit量化只能表示16个离散值，1-bit只有2个，三元量化（-1/0/+1）只有3个可能值65% similarUnverified千问3.8 27B基于千问3.5的整套架构体系，架构层面与3.5完全一致，仅做了继续预训练和后训练（含SFT与RL强化）62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/795640API
curl https://kongchang.com/api/v1/knowledge/claims/795640MCP
get_claim(id=795640)