X · X// 3h agoBENCHMARK RESULT

DigitalOcean tops Artificial Analysis benchmarks

DigitalOcean is promoting its Serverless Inference offering after benchmarking DeepSeek V3.2, MiniMax-M2.5, and Qwen 3.5 397B against other providers. The company says its DeepSeek V3.2 setup reaches 230 output tokens per second with sub-1-second TTFT on 10,000 input tokens, and that its results place it near the top of the April 2026 Artificial Analysis leaderboard. The post frames the gains as the result of co-designing the stack with Inferact, tuning vLLM, and optimizing the full serving path on NVIDIA Blackwell Ultra hardware.

// ANALYSIS

Strong infrastructure benchmark post: this is DigitalOcean selling speed as a product feature, not just raw cloud capacity.

–The core claim is performance, not model novelty: faster TTFT and higher output tokens/sec for three frontier open-weight models.
–The message is credible enough to matter because it cites Artificial Analysis and gives concrete numbers, but it is still a vendor-published benchmark, so workload parity and methodology details matter.
–The engineering angle is the real story: hardware selection, speculative decoding, quantization, and serving-stack tuning are presented as the reason the numbers are competitive.
–If these results hold in production, this is a meaningful signal for teams choosing an inference provider based on latency-sensitive user experiences.

// TAGS

digitaloceanserverless inferenceinferencebenchmarkingartificial analysisdeepseekminimaxqwenvllmai infrastructure

DISCOVERED

3h ago

2026-04-30

PUBLISHED

3h ago

2026-04-30

RELEVANCE

9/ 10

AUTHOR

digitalocean