Frontier models on the GPUs
you can actually get.
GraphX runs open models like DeepSeek, GLM and Kimi on commodity, in-stock silicon, at a fraction of the price of the data-centre hardware they were built for. Pick a model, and we find where it can run today.
DeepSeek V4 Flash on eight NVIDIA L4s
A 284B-parameter model whose reference deployment is an 8× B200 node, served by GraphX on 8 low-power inference cards with 192 GB of memory between them. The native 146 GB of weights would leave almost no room for conversations; GraphX packs them into about 124 GB.
| GraphX on 8× L4 | Reference: 8× B200 | |
|---|---|---|
| GPU memory | 192 GB | 1,536 GB |
| Memory bandwidth | 2.4 TB/s | 64 TB/s |
| GPU power | 576 W | 8,000 W |
| Price per hour | $13.35 | $113.93 |
| Cards to buy | $20,800 | $360,000 |
| Weights in memory | ~124 GB (GraphX, ~3.5 bits) | 146 GB (FP4 + FP8) |
Measured with GraphX Performance Monitor: realistic simulated users with 2,000 to 4,000-token documents, Oct 25, 2026 and earlier. Prices are AWS on-demand list prices for both nodes; card prices are approximate street prices. GraphX serves the model at about 3.5 bits per weight; a published accuracy comparison at that precision does not exist yet.
Stop waiting for B200s.
The reference architectures for today's frontier models assume the scarcest, most expensive accelerators on the market. GraphX changes what the model needs.
Runs on silicon in stock
L4, L40S and RTX PRO 6000 are rentable in most regions today. GraphX fits models that were published for B200 and H200 nodes into them.
A fraction of the price
A smaller, cheaper node at list price, often from a neocloud, instead of a flagship node at a hyperscaler. We show the price of every option before you deploy.
Measured, not promised
Every figure on this page comes from load tests we ran against live GraphX deployments with simulated users, and every reference figure cites its source.
Model in, endpoint out.
Pick a model
See its size, its storage on disk, the GPU memory it normally needs and the reference hardware it was published for.
We scan the market
Hyperscalers and neoclouds are checked for every GPU configuration GraphX can run the model on, with price and stock for each.
Deploy in one click
Choose the provider and configuration. GraphX is deployed and serves an OpenAI-compatible endpoint.