Summary
According to a post on X by @ArtificialAnlys, DeepSeek V4.1 Flash speeds on GB300 NVL72 rose up to about 1.5× tok/s per agent over the past month on AA-AgentPerf. Single-agent speed went from 596 to 848 tok/s (about 1.42×), and throughput per GPU reached about 1.6× at roughly 80 tok/s per agent. The post credits newer vLLM kernels, FlashInfer autotuning, CUDA graphs, speculative decoding, and NVIDIA Dynamo with vLLM prefill/decode running on the same GPUs.
AI-assisted summary based on the listed source.
What happened
On AA-AgentPerf, Artificial Analysis says DeepSeek V4.1 Flash speeds on GB300 NVL72 rose up to about 1.5× tok/s per agent over the past month. Single-agent speed is 848 tok/s vs 596 before (~1.42×), with ~1.6× throughput per GPU at ~80 tok/s per agent, credited to newer vLLM kernels, FlashInfer autotuning, CUDA...
Why it matters
For teams serving open-source models like DeepSeek on NVIDIA hardware, these figures suggest that inference software updates alone can change speed on existing GPUs. Builders and operators comparing hosting costs should watch whether these gains appear in other benchmarks and real workloads. The source snippet is cut off, so the full reasoning in the post is not visible here.
What this means for you
Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.
Signal Intelligence
Signal Strength 85%
Technical label WATCH
Public Interest 56
Category ROBOTS & HARDWARE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 82
Practical Impact Score 26
Novelty Interest Score 94
Consequence Score 18
Curiosity Score 16
Shareability Score 68
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
