Unverified50% confidenceOpinionExact time
Multi-View MoE has so far been validated only at the 10B-A1B scale, and whether the gains hold for larger models and larger numbers of routers remains to be seen.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/30/2026
First Seen
Valid until: 1/28/2027
Sources
Related Entities
Related Claims
UnverifiedAfter training a 24-layer, 10B-A1B model on 708B tokens, the full Multi-View MoE improves the average score on nine downstream tasks by 1.18 percentage points over the baseline.73% similarUnverifiedIn MoE models, the weights of all experts still need to be loaded into VRAM on standby, so despite the computation being close to that of a 49B dense model, the VRAM footprint approaches the scale of the full 1T model.64% similarUnverifiedThe VRAM usage of MoE models is still proportional to the total number of parameters; although computation is fast, the model file itself remains large.63% similarUnverifiedMoE models activate only approximately 10 out of 512 experts per token, with an effective activated parameter count of approximately 20–30B.60% similarUnverifiedThe scale at which Reviser has currently been validated remains at the hundred-million parameter level, and whether it can scale to mainstream large model sizes remains to be seen.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115296API
curl https://kongchang.com/api/v1/knowledge/claims/1115296MCP
get_claim(id=1115296)