In the news
How we trained the fastest DSpark for Kimi-K3 using GB300 NVL72
vLLM · Helen Zhao, Fynn Schmitt-Ulms, Yuchen Fama, Antonio J. Dominguez, and Kevin Li · Published · 3 min read
In 30 seconds
- What happened
- vLLM's Speculators library trained a DSpark speculative decoding model for Kimi K3, boosting single-stream interactivity from 110 to 435 tokens per second.
- Why it matters
- Matters for engineers deploying large language models who need faster response times and higher throughput without sacrificing latency under concurrent load.
- Watch out
- DSpark requires careful hardware configuration and disaggregated multi-node training. Performance gains vary significantly by workload type and request concurrency levels.
- kimi
The patterns behind this
- Speculative & Parallel Tool Execution
- Agentic Context Engineering (Evolving Playbook)
- Skill Library (Voyager)
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.