Unverified50% confidenceFactExact time
Strata allocates approximately 6 GB of VRAM as an expert cache storing around 4,500 experts, while all 24,576 experts are pinned in system memory.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/5/2026
First Seen
Valid until: 1/3/2027
Sources
Related Entities
Related Claims
UnverifiedStrata converts the entire computer into a three-tier memory hierarchy: VRAM holds what each token touches, system memory pins all approximately 24,000 groups of experts, and the SSD stores a 29 GB N-gram lookup table.77% similarUnverifiedStrata uses low-bit quantization; the minimum configuration requires approximately 38 GB of combined memory and VRAM, the highest-quality configuration requires approximately 55 GB, and 64 GB of system memory is the ideal target.67% similarUnverifiedStrata additionally downloads a data directory of approximately 6 GB or more, which contains MTP (multi-token prediction) related content.67% similarUnverifiedWhen the context exceeds 64,000 tokens, Strata keeps most of the KV cache in memory and retains only the hottest 32,000 tokens in VRAM, improving speed from approximately 51 to 63 tokens/second at a 262,000-token context.66% similarUnverifiedStrata adopts hierarchical KV cache management, keeping only 32K units of KV cache in VRAM while offloading the rest to memory; in actual testing the KV VRAM hit rate reached 98.9%.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/978805API
curl https://kongchang.com/api/v1/knowledge/claims/978805MCP
get_claim(id=978805)