Unverified50% confidenceFactExact time
The workload measured here is the public typed-decisions row set: 400 cases and 2000 decisions, with the local model being Phocinae-Largha-150M-v1 (CPU, fp32) and the hosted model used for comparison being deepseek-flash in the flash tier (temperature 0).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedThe test uses a machine equipped with 96GB of VRAM and 188GB of system memory, achieving hybrid inference by offloading the embedding table to system memory.65% similarUnverifiedThe project adopts a mixed-precision quantization strategy: 8-bit for dense layers, 4-bit for expert layers, and 8-bit quantization for the KV cache.64% similarUnverifiedTaking the LLaMA-3 70B model as an example, with 80 layers, head_dim=128, and BF16 precision, each token occupies approximately 80KB of KV cache.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1016997API
curl https://kongchang.com/api/v1/knowledge/claims/1016997MCP
get_claim(id=1016997)