A fit for engineering teams that want a prepaid endpoint for DeepSeek Flash plus an Ideogram image API, with a compiler-first route to on-prem later. It is early: one live text model, no published SLA, and a team of about ten or fewer.
What changed
- · Release Pay-as-you-go API lists DeepSeek V4.1 Flash text and Ideogram 4.5 images source
Luminal is a San Francisco inference company that compiles AI models into GPU and ASIC code, sells serverless per-token text and per-image API access, and licenses on-prem deployments. Its compiler is open source under dual licences. The vendor reports 3.2x the throughput of vLLM on GPT-OSS 120B across eight H100 GPUs.
Frequently Asked Questions
What does Luminal charge for inference?
Text on the live DeepSeek Flash model is billed per million tokens: $0.1425 input, $0.0029 for cache-hit input and $0.5700 output, reasoning included. Images cost $0.0080 to $0.2000 each depending on the quality tier. Everything is drawn from a prepaid balance with a $50.00 minimum top-up, and there is no subscription. On-prem licensing is quoted by sales.
Can you rent GPUs directly from Luminal?
No GPU rental plan is listed. Luminal sells token and image APIs plus licensed on-prem deployments. Its benchmark reports name the data-center GPUs the compiler was tested on, but those are test rigs, not instances available to buy.
Which providers compete with Luminal?
Groq pairs custom chips with open-model serving, Fireworks AI hosts a broader open-model menu, and Baseten suits teams deploying their own weights with SOC 2 Type II. Choose Luminal when compiler-level optimization or a later on-prem move matters more than catalog size.
Is Luminal a better pick than a model router for open models?
They solve different problems. A router like OpenRouter gives access to many models through one key, while Luminal runs a narrow set on its own compiled stack. Teams committed to its hosted DeepSeek model can test the endpoint behind a router and compare latency and cost before switching traffic.
How do you get started with Luminal?
Sign up at dash.luminal.com, add prepaid credit by card, and create an API key. The same key works on the image host at images.luminal.cloud. Default limits are 60 requests per minute, and higher limits are requested by emailing the founders with expected load.
Top Alternatives
- Baseten: Pick Baseten for custom weights and certified deployments; pick Luminal for a prepaid hosted DeepSeek endpoint.
- Atlas Cloud: Pick Atlas Cloud for GPU rental and a broad model list; pick Luminal for compiler-level optimization.
- CoreWeave: Pick CoreWeave for raw large-scale GPU capacity; pick Luminal when you want inference delivered, not capacity.