Unverified50% confidenceFactExact time
Betelgeuse's 110 GB Engram table is placed entirely on the NVMe SSD, is activated only in layers 1 and 14, touches only 48 rows (about 192 KB) per token, and the active rows are staged into a 2MB VRAM buffer at t=0.
1
Sources
50%
Confidence
Long-term
Relevance
11/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe Ngram lookup table only reads a few rows per token and does not need to reside permanently in high-speed memory; after stripping out this table, the 2-bit build of the core model requires only 152GB, leaving nearly 100GB of headroom on a 256GB machine.73% similarUnverifiedThe 28.8 GB N-gram interpolation table resides on the hard drive, occupying neither VRAM nor memory.72% similarUnverifiedIn Unsloth, keeping the Engram table of DeepSeek-V4.1-Flash-1017 on the SSD, the model takes about 125 GB of memory to run under 1-bit quantization and about 300 GB under 4-bit quantization.72% similarUnverifiedEvery 1,000 in-flight tokens require about 0.33 GB of VRAM in the KV cache.71% similarUnverifiedApproximately 200 GB of n-gram tables are stored in regular system memory rather than on the GPU, with background transfer prefetching the rows about to be used.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1147046API
curl https://kongchang.com/api/v1/knowledge/claims/1147046MCP
get_claim(id=1147046)