Run any AI model the fastest, cheapest way, on any hardware. Tell Ceptron the model, the hardware you have or could rent, and what you care about. It tells you the setup to use and the speed and cost to expect — in seconds, for models and hardware it has never seen.
For teams serving open models: AI startups, inference providers and GPU clouds.
Every model can run dozens of ways: how many GPUs, which GPUs, what precision, how many users per batch, what to keep in memory. The right answer depends on the model, the hardware and your traffic.
Teams benchmark setup after setup. Vendor sizing tools need hours of profiling per GPU type, and cover one vendor.
Sensible-looking setups for the same model can differ by 3× or more in cost per token.
A new model, GPU or engine version every few weeks makes last month's benchmark obsolete.
Choose a model and what you care about, tick the hardware you could use, and see the setups ranked. Click a row to see where the model lives in memory and what limits its speed. Every answer was computed by the Ceptron physics engine from the model's architecture and the hardware's datasheet; nothing is written by hand.
The demo shows the upper limit each setup can reach, from the model's architecture and the hardware's datasheet. Real serving engines reach a fraction of that limit, and the fraction depends on the engine, its version and the machine. That is what Ceptron learns, so its answers reflect your stack, not just the datasheet. Give it a 32 GB laptop and GLM-5.3-Flash and it streams experts from disk; give it H200s and it keeps everything in GPU memory: the same strategies the Colibri engine and the official vLLM recipe use.
Model facts come from each public checkpoint and its model card. Hardware facts come from vendor datasheets; specs marked * are unverified estimates. Prices are rough on-demand rental rates per GPU-hour, and local machines are purchase price spread over 3 years of 24/7 use. All are assumptions, not quotes. Every speed is a physics limit, not a measured result.
Point Ceptron at a public checkpoint and it maps the architecture without downloading the weights. Here is GLM-5.3-Flash: every number below matches its model card and the official vLLM recipe.
The forecast, the recommended config, and how sure it is.
An optional 5-minute run on your own machine, for tighter answers.
Work you do not redo: shared prompt prefixes and experts already in fast memory.
The cost of moving data between GPUs for each kind of parallelism, on the actual topology.
Everything you can set. It is the search space Recommend explores, and every knob is a mechanism in the forward model, so the world model can reason about it.
When actual differs from planned, find out why, and fix the world model rather than the number.
Turns seconds into dollars and joules.
Given every parameter, it computes the outcome. It steps through a serving engine's life, so batching, caching and queueing come out of it rather than being assumed.
Every NVIDIA and AMD part from 2016 on, normalized so the physics can use it: compute per number format, memory capacity and bandwidth, caches, interconnect, power.
Decides where every byte lives (GPU memory, host RAM or NVMe) and what that leaves for the KV cache.
The model's architecture, read straight from its checkpoint: which weights exist, their shapes and number formats, and which of them each token uses.
The forward model plus a range for everything unknown. Each sample is one complete world.
Arrivals against service rate: where tail latency comes from.
Each operation takes as long as whichever runs out first: compute or memory bandwidth.
Turns per-step time into per-request time, the way a serving engine actually batches.
What you tell it: the model, the hardware you have or could rent, the engine, and what you care about.
Drafting several tokens and verifying them in one pass. The output is identical; only speed changes.
Which engine, version and kernels run the model. The biggest source of error: the same GPU reaches about half or most of its peak depending on software.
One network learns, across all those worlds, how serving setups behave. Paid once, on our GPUs.
The trained network. Every question is one forward pass, answered in seconds, for setups it has never seen.
What you send the server. The same server is fast or slow depending on lengths, burstiness and how much prompts share.
Millions of samples from the prior. Each one is a different serving setup and how it performs.
Tell us which models and hardware you serve and we will be in touch.