Raw SignalOct 25, 20261 min read5 sources

Who serves an open model now matters as much as which model you pick

Vals AI found cost differences of up to 2.3× between providers serving the same open-weight model on agent tasks, and the model makers' own APIs were neither the cheapest nor the fastest.

In one minute

3× between providers serving the same open-weight model on agent tasks, and the model makers' own APIs were neither the cheapest nor the fastest. Google Cloud reports up to 5x higher performance for Beam depending on the serving stack, and vLLM shows rack-scale tuning for DeepSeek.

Vals AI found cost differences of up to 2.3× between providers serving the same open-weight model on agent tasks, and the model makers' own APIs were neither the cheapest nor the fastest. Google Cloud reports up to 5x higher performance for Beam depending on the serving stack, and vLLM shows rack-scale tuning for DeepSeek. NerfBench now tracks whether a served model keeps performing as it did. That makes the endpoint something to benchmark and keep checking.

From the Field Notes:

Evidence / sources

  1. 1vals.ai/benchmarks/inference_provider
  2. 2cloud.google.com/blog/topics/ai-infrastructure/day-0-support-with-5x-perf-for-reflection-beam-on-google-cloud
  3. 3vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep
  4. 4github.com/vllm-project/vllm/pull/62009
  5. 5bridgebench.ai/blog/the-methodology-of-nerfbench
All Raw Signals

Next reads

Read next

Open the graph
  1. 01topicGoogleExplore the google topic hub.
  2. 02archiveRaw SignalsOpen every source-linked raw signal.
  3. 03source ledgerSource LedgerInspect the public source evidence graph.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract