Unverified50% confidenceFactExact time
The Kirin 9050 series adopts a multi-level compute-storage coordination scheme, loading hot experts involved in computation into DDR and keeping cold experts in Flash, and together with low-bit quantization it can run a 30B MoE large model while using less than 2.5GB of additional memory.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
Sources
Related Entities
Related Claims
UnverifiedThe 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).66% similarUnverifiedA 125-billion-parameter model can run on 12G of VRAM, relying mainly on model quantization and a mixture-of-experts (MoE) architecture, achieved by activating only part of the expert parameters, quantizing weights to 4bit or lower, and dynamically loading inactive parameters from memory or SSD on demand.64% similarUnverified70B-scale models typically require multi-GPU setups or cloud-based A100/H100s for smooth inference.63% similarVerifiedEven after quantization, 70B-level models often require more than 40 GB of VRAM or a multi-GPU configuration.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1033112API
curl https://kongchang.com/api/v1/knowledge/claims/1033112MCP
get_claim(id=1033112)