Inferact Platform

Inferact Platform is a commercial production inference service built on the open-source vLLM engine by the team that created it. It is aimed at AI platform and infrastructure teams that serve open-weight models at scale and want vLLM performance with autoscaling, metering, observability and support, without building a dedicated inference team.

Inferact is the company that creators and core maintainers of vLLM, the Apache 2.0 inference engine that began in UC Berkeley's Sky Computing Lab in 2023, launched in January 2026 with a $150 million seed round at an $800 million valuation, co-led by Andreessen Horowitz and Lightspeed. Simon Mo, one of vLLM's original creators, is CEO, and Databricks co-founder Ion Stoica is on the founding team. The Inferact Platform is the paid product. Customers either deploy it on their own cloud or datacenter (bring-your-own-cloud, with an Inferact control plane for monitoring and rollouts) or use Inferact Cloud, where Inferact provisions and runs the GPUs. The platform adds autoscaling, day-zero support for new open models, custom kernels and trained speculative decoders sold as Frontier Tuning, authenticated endpoints with rate limits and token metering, and dashboards for throughput, latency and cost. Forward-deployed inference engineers work with customer teams from pilot to production. Inferact says the platform is SOC 2 compliant and supports data residency controls. Its site cites 18.3 million tokens per dollar on SemiAnalysis AgentX with Kimi K3 and a first-place ranking on Artificial Analysis in May 2026. In September 2026 Google Cloud and Inferact announced an engineering partnership to make TPUs a first-class vLLM target, and on October 15, 2026 Inferact announced a collaboration with AMD covering vLLM on AMD Instinct GPUs. Inferact does not publish pricing. It competes with SGLang's commercial arm RadixArk, NVIDIA TensorRT-LLM and hosted inference clouds such as Fireworks AI and Baseten.

Category

Infrastructure & Cloud

Pricing

Contact for pricing

Target Market

CTOs, Heads of AI Infrastructure, ML Platform Engineers, AI Labs

Deployment

Hybrid, Open-source, Multi-cloud

• ✓
  Bring-your-own-cloud deployment

  Runs the platform in your own cloud or datacenter with an Inferact control plane for monitoring and rollouts, so model traffic stays in your account.

• ✓
  Inferact Cloud

  Inferact provisions and operates the GPUs for teams that want managed endpoints and do not want to buy or reserve capacity themselves.

• ✓
  Frontier Tuning

  Custom kernels and trained speculative decoders built per model, aimed at lowering cost per token and raising tokens per second on production workloads.

• ✓
  Metered, authenticated endpoints

  Endpoints ship with authentication, rate limits and token metering, which lets platform teams charge back usage to internal consumers.

• ✓
  Day-zero model support

  New open-weight model releases are supported at launch through vLLM, including Kimi K3, which Inferact says passes the Kimi Vendor Verifier across all benchmarks.

• ✓
  Inference observability

  Dashboards track throughput, latency and cost per model so operators can see where capacity and spend go.

• ✓
  Forward-deployed engineers

  Inferact inference engineers work alongside the customer's team from pilot through production rollout and tuning.

• •
  Self-hosted open-model serving

  Serve a large mixture-of-experts open model on reserved GPUs inside your VPC, with autoscaling and token metering handled by the platform.

• •
  Agentic workload inference

  Run coding or research agents that resend long shared contexts, using prefix caching and speculative decoding to cut latency and cost per call.

• •
  Multi-accelerator inference

  Move serving between NVIDIA GPUs, AMD Instinct GPUs and Google TPUs on the same vLLM-based stack to follow capacity and price.

• •
  Internal inference platform

  Give product teams authenticated, rate-limited endpoints with per-team token metering so central IT can allocate GPU cost by consumer.

Inferact publishes no list prices; the site offers only 'talk to an engineer' and meeting booking. The underlying vLLM engine stays free under Apache 2.0, so the paid product is the managed control plane, Inferact-run GPUs, per-model tuning and engineering support. Token metering exists on endpoints, but the commercial meter (per token, per GPU-hour or contract) is not disclosed.

This page was written from 10 sources, 7 on domains other than inferact.ai.

1. 1.inferact.ai — inferact.aivendor
2. 2.inferact.ai — inferact collaborates with amd to expand agentic inference tvendor
3. 3.inferact.ai — google tpu partnershipvendor
4. 4.siliconangle.com — inferact launches 150m funding commercialize vllm
5. 5.techcrunch.com — inference startup inferact lands 150m to commercialize vllm
6. 6.pulse2.com — inferact launches with 150 million funding at 800m valuation
7. 7.github.com — vllm
8. 8.docs.vllm.ai — latest
9. 9.premai.io — 10 best vllm alternatives for llm inference in production 20
10. 10.hn.algolia.com — search