vLLM on AgentX: up to 106× serving-cost advantage over Opus 5 API pricing
vllm_project · x · 2026-09-09
The vLLM team published a deep dive on optimizing serving for real-world agentic workloads, now a major share of vLLM traffic with multi-turn sessions, long contexts and heavy prefix reuse.
- On SemiAnalysis's public AgentX benchmark, vLLM hits up to 130K tokens per GPU-second on DeepSeek V4 Pro and 376 tokens/s interactivity on MiniMax M3.
- Across DeepSeek V4 Pro, MiniMax M3 and Kimi K3, vLLM delivers a 14.6×–106× serving-cost advantage over Opus 5 API pricing.
- Hard-won lessons: pipeline parallelism wins on cold long prompts but bubbles eat gains on warm agent turns; decode context parallelism matched DEP on DeepSeek V4 but won on Kimi K3—parallelism must follow the attention stack; load balancing doesn't always beat session-sticky routing.
- Everything is reproducible via a public dashboard with configs across GB300 NVL72 and B300.
More from Infra
- Even an OpenAI bankruptcy wouldn't stop the AI buildout — the GPUs just change hands — markjeffrey · 2026-09-09
- llama.cpp on 2x RTX 2060 hits 45 tok/s on Qwen3.8 27B — is dual RX 6800 worth it? — BarberIcy366 · 2026-09-09
- Baseten tops Coval's voice AI benchmark: STT ~5x faster than OpenAI with lowest WER — baseten · 2026-09-09
- 2×4090 llama.cpp concurrency: soft cap of 5 agents at 64k context, hard cap 9 — three weeks of data — Iamisseibelial · 2026-09-09
- exe.dev deep dive: ssh to a persistent Linux VM in half a second, priced like a folder — davidcrawshaw · 2026-09-09
- GPT-6 Astra lands on Amazon Bedrock with 1M-token context and first Critical cyber rating — AWS ML Blog · 2026-09-09