Verda has launched a preview inference service for AI teams, starting with Moonshot AI’s Kimi K3.1 Pro. The pitch is not just another place to run a model: Verda says its European data centres and caching are built to keep long-running AI agents responsive and reduce repeated processing.
Verda Watch analysis
What happened
Verda Inference provides managed access to open-weight models, with Verda handling the GPUs, models and serving. The preview starts with Kimi K3.1 Pro, which Verda describes as a 3-trillion-parameter-class model with a one-million-token context window. Verda says it will add new Kimi models as their open weights are released.
The service uses an OpenAI-compatible API, so developers can switch by changing the base URL, API key and model ID. Verda says customers pay for tokens used, with cached input priced lower, and can manage inference alongside their other Verda compute services. Usage limits apply during the preview.
Why it matters
Long-context agents repeatedly send much of the same material as they work. Verda says its shared cache reuses about 99% of repeated context in testing, while moving sessions that exceed GPU memory into a larger memory tier rather than recomputing them from scratch. It also separates prompt processing from answer generation across GPUs to prevent long prompts from holding up responses.
Verda claims up to 10 times faster time to first token on long-context agent workloads than standard caching. That comparison comes from its own internal test with Kimi K3.1 Pro and a one-million-token coding-agent context, not an independent benchmark. The company also says it fit 16 times more agent sessions on each GPU in six weeks, measured on SemiAnalysis’s AgentX workload.
Our read
This is a substantial infrastructure launch because it addresses a real nuisance of agentic work: each step can drag a growing history behind it. If Verda’s caching and separation of workloads hold up beyond its own tests, developers could spend less time waiting and less money reprocessing old context. The compatibility path is refreshingly practical too. The figures are promising, but they are company-reported; builders will want to test them against their own workloads before moving anything important.
What to watch
- Which open-weight models Verda adds, and how quickly they arrive.
- Whether independent or customer testing supports the company’s speed, caching and capacity claims.
- How preview usage limits and token pricing affect longer workloads.
- Whether developers can switch existing tools with only the stated API changes.
Discussion spark: For long-running AI agents, would you prioritise faster responses and lower repeated-token costs, or keep workloads with the model provider that already serves them?
Sources and evidence
- Source update (4 November 2026, 00:00 UTC)
not affiliated with or endorsed by Verda