Unverified50% confidenceFactExact time
Alibaba's Qwen 3.8 Flash Next uses an MoE architecture with a total of 125 billion parameters, activating only 6 billion parameters per token output, and places a 51-billion-parameter common phrase lookup table in regular RAM so it does not occupy VRAM.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/22/2026
First Seen
Valid until: 12/21/2026
Sources
Related Entities
Related Claims
UnverifiedMoE models read only the activated subset of parameters at each step, making them much faster than dense models of the same parameter count; for example, GLM 5.3 Flash reads only about 30 GB per token.66% similarUnverifiedCombining MoE with quantization technology can make it possible for a laptop with 16 GB of memory to run MoE models with tens of billions of parameters.66% similarUnverifiedMoE模型采用稀疏激活，Qwen 3.5 122B总参数122B但每次推理仅激活约10B，需将全部参数加载进内存，对内存容量极度敏感64% similarUnverifiedFor a typical 70-billion-parameter model, the KV cache occupies approximately 0.5 MB of memory per token, and generating 100,000 tokens requires carrying 50 GB of cache.63% similarVerifiedA single inference at the million-token level requires tens to hundreds of gigabytes of high-end GPU memory (A100/H100 class).61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/941012API
curl https://kongchang.com/api/v1/knowledge/claims/941012MCP
get_claim(id=941012)