Unverified50% confidencePredictionExact time
The author predicts that if the Qwen team releases a Flash version optimized for TP4, inference speed on DGX Spark could be boosted to the 100 tokens/s level.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUnder a dual-Spark TP2 configuration, Qwen 3.8 Flash achieves an inference speed of 55–70 tokens/s and supports 4 fully concurrent agents running simultaneously.75% similarUnverifiedFor users with 4×DGX Spark, a tiered deployment is recommended: the supervisor agent runs GLM 5.3 Flash on TP4, worker agents run Qwen 27B, and sub-agents run Qwen 3.6 35B.70% similarPartially VerifiedQwen3.8-Flash-Next adopts a Mixture of Experts (MoE) architecture with 125B total parameters and 6B activated parameters.67% similarUnverifiedAlibaba's Tongyi Qianwen officially released the Qwen3.8-Flash-Next (Qianwen 308 Flash NEX) model, with a total parameter count of 181.5B.66% similarUnverifiedIn the official demo, PAIR can assign the Qwen model to an RTX 5090, assign the Qwen MoE model to a DGX Spark, and work in conjunction with Radeon graphics cards.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/904313API
curl https://kongchang.com/api/v1/knowledge/claims/904313MCP
get_claim(id=904313)