Unverified50% confidenceFactExact time
Pathway推出的BDH/CQ系列模型只有1.5亿参数，采用后Transformer架构，在循环隐空间中持续更新记忆而非把推理过程写成token，获得5亿美元估值
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/22/2026
First Seen
Valid until: 11/20/2026
Sources
Related Entities
Related Claims
VerifiedThe autoregressive generation process of reasoning models is strictly sequential; KV Cache VRAM usage grows linearly with the length of the thinking token sequence, making memory bandwidth the performance bottleneck.66% similarUnverifiedTransformer-based LLMs perform an independent forward-pass computation for each inference, with model weights fixed during the inference phase; what is referred to as context memory is essentially the re-injection of historical conversation text into the prompt.63% similarUnverifiedIn standard Transformer inference, generating each new token requires accessing the key-value vectors of all historical tokens, and the memory footprint of the KV cache grows linearly with sequence length.60% similarUnverified2024年中期推理模型首次上线时用户反馈简单问题等待过久，催生了自动路由机制60% similarUnverifiedThe Transformer architecture trades native sequential memory for parallel training capability, whereas RNNs/LSTMs pass state between time steps via hidden vectors and naturally possess sequential memory.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/786987API
curl https://kongchang.com/api/v1/knowledge/claims/786987MCP
get_claim(id=786987)