Unverified50% confidenceFactExact time
Beam's activation ratio is about 4.6%, lower than Mixtral 8x7B's roughly 46%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
Sources
Related Entities
Related Claims
UnverifiedBeam's total parameter count is 501B, with 23B parameters activated per token (about 4.6% of the network is activated).63% similarUnverifiedMixtral 8x7B的实际激活参数量仅为总参数量的约1/460% similarUnverifiedReflection claims Beam's reasoning ability is comparable to GLM-5.2, but its reasoning compute is about 1/3 to 1/4 of the latter's.59% similarUnverifiedBeam adopts a Sparse Mixture-of-Experts (Sparse MoE) architecture with 52 layers, activating only 23 billion parameters at a time.57% similarUnverifiedW4A4量化将权重和激活值均压缩至4位整数，理论上相比FP16可将显存占用降低至约1/455% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1036253API
curl https://kongchang.com/api/v1/knowledge/claims/1036253MCP
get_claim(id=1036253)