Unverified50% confidenceFactExact time
DeepInfer sets three expectations for hardware: expert weights in FP4, activations and KV in FP8, and scaling factors in the low-precision UE8M0 format; KV Cache per token under 1KB; and the bottleneck shifting to large-scale high-speed interconnect and device-side communication.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/25/2026
First Seen
Valid until: 1/23/2027
Sources
Related Entities
Related Claims
UnverifiedDeepSeek already proposed during the V3.1 period that UE8M0 FP8 was designed for next-generation domestic chips; this time it is fully implemented in DeepInfer, meaning chips that do not support this format will directly suffer in precision or performance.79% similarUnverifiedWhen DeepInfer was designed, three goals were set: be compatible with chips from different vendors, organize prefill and decode around common hardware, and keep the KV cache for at least 72 hours.76% similarUnverifiedDeepSeek-V4.1 uses an asymmetric activated-parameter design: only 8B parameters are activated on input and 16B parameters on output, while KV Cache and memory usage are significantly reduced compared with the previous generation.73% similarUnverifiedThe interconnect bandwidth of DeepSeek operators on Huawei Ascend supernodes has approached the hardware limit.72% similarUnverifiedThere is currently no public information on which hardware DeepInfer supports, whether it is compatible with other models, or how it actually performs compared with mainstream inference frameworks.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1095890API
curl https://kongchang.com/api/v1/knowledge/claims/1095890MCP
get_claim(id=1095890)