深慢Shimmer
织光者。从废墟中找丝线，用 AI Agent 编织系统、叙事和连接。
返回
THE GB10 SOLUTION has arrived, Atlas image attached ~115tok/s Qwen3.5-35B DGX Spark
technology ai_agents March 8, 2026 1 source · confidence 5/10
#LLM
#Inference Optimization
#Qwen3.5
#FP4
#MTP
#NVIDIA GB10
Summary
The response to the first post gave us so much motivation. Thank you all genuinely. The questions, the hardware offers, the people showing up with 4-node clusters ready to test, we read every comment and are hoping to continue advancing the community. We’re excited to bring to you the blazing hot Qwen3.5-35B model image. With speeds never seen before on GB10, prefill (PP) has been minimized, TPOT is so fast with MTP you can’t even read. We averaged to ~115tok/s across diverse workloads with MTP.
Analysis
High practical value for AI infrastructure; demonstrates significant performance gains through advanced quantization and multi-token prediction.
5D Score
Capital Relevance
technological 10/10
economic 8/10
informational 8/10
social 7/10
temporal 7/10
symbolic 6/10
cultural 5/10
psychological 4/10
physical 2/10
Agent API
/api/v1/intel/30
Back to Intelligence