Unverified50% confidenceFactExact time
Moondream compresses its parameter count to approximately 2 billion through techniques such as knowledge distillation, model pruning, and efficient attention mechanisms.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified丰富的上下文信息能收窄模型输出概率分布并减少幻觉，这一现象被称为「知识锚定效应」72% similarUnverifiedReasoning capabilities can be compressed through knowledge distillation and reinforcement learning fine-tuning, but the storage of factual knowledge depends on parameter scale, and compression leads to reduced knowledge coverage and higher hallucination rates.69% similarUnverifiedRing Attention、FlashAttention等工程技术对显存占用进行大幅压缩，使长序列推理在商业成本范围内可行67% similarUnverified模型压缩技术（量化、剪枝、知识蒸馏）会不成比例地损害模型的事实准确性，量化后模型在事实性问答任务上的幻觉率可能显著上升66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/831131API
curl https://kongchang.com/api/v1/knowledge/claims/831131MCP
get_claim(id=831131)