Unverified50% confidenceBenchmarkExact time
Models using MLA such as DeepSeek-V4.1-Flash compress the KV state to about 2.4 GB per 32k stream (FP8 KV), while wider architectures consume 3.0 GB–3.8 GB per 32k stream.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/14/2026
First Seen
Valid until: 1/12/2027
Sources
Related Entities
Related Claims
UnverifiedUnder a configuration using INT4 dynamic quantized weights and FP8 for KVCache, the weights occupy 21.9GB of VRAM, KV occupies 8.5GB, and the total KVCache capacity is 850K tokens.77% similarUnverifiedWhen the context of LLaMA-2 7B is extended to 32K tokens, KV Cache GPU memory usage expands to approximately 16 GB, approaching the total memory of a single consumer-grade GPU.76% similarUnverifiedAfter compression, the DeepSeek V4 Flash model size was reduced from nearly 500 GB to 80.8 GB.75% similarUnverifiedQ3_K_M compresses each weight to approximately 3 bits, reducing the weight size of a 27B model to approximately 12–13 GB.74% similarUnverifiedIn the DeepSeek V1 era, each incremental token required nearly 400 KB of KV cache, which dropped to 48 KB in V3.2.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027760API
curl https://kongchang.com/api/v1/knowledge/claims/1027760MCP
get_claim(id=1027760)