Serve Step 5's long sessions on 45% fewer GPUs.
Step 5 moves about 14 trillion tokens a week on OpenRouter (as of Oct 18, 2026), and most of that traffic is long, cached agent sessions. HyperCache shrinks the memory each session holds, so the GPUs you already run carry about 1.8 times as many of them.
Estimate. GPU spend of $80M to $240M a year for 14T tokens a week of Step 5, with 15 to 35% saved. Assumes about $2 per GPU-hour.
More sessions on the GPUs you already run.
1.8× more concurrent long sessions
Each 100k-token Step 5 session takes a little over half the cache memory it takes in FP8. The same node holds about 1.8 times as many open conversations before it has to queue or evict.
More paused agents stay warm
Step 5 traffic is mostly cached input. Smaller sessions mean far more of them stay resident between tool calls, so fewer returning requests pay for a full prefill. The cache-hit rate goes up.
Longer contexts at the same cost
A 256k session fits in about the memory a 140k session needs today, so you can offer the full context window without reserving extra GPUs for it.
Lower cost per token
On memory-bound long-context traffic, fitting 1.8× the sessions per GPU cuts GPU-hours per token by up to 45%. Price below the market and keep your margin, or hold price and widen it.
Ready on new Step releases
Our calibrator tunes HyperCache to each new Step checkpoint and checks quality at 100k+ context before you serve it. Every bundle ships with its own quality report.
Up to $85M a year across all Step 5 traffic.
| Weekly Step 5 tokens | Share of OpenRouter volume | Saved per year |
|---|---|---|
| 0.5T | about 4% | $0.4M to $3.0M |
| 1T | about 7% | $0.9M to $6.1M |
| 5T | about 36% | $4.3M to $30M |
| 14T | all of it | $12M to $85M |
Basis: 14T tokens a week is about 728T tokens a year, roughly $160M to $480M a year of serving revenue depending on cache-hit rate and price, of which about half is GPU cost. The saving is 15 to 35% of that GPU cost, depending on how much of your traffic is long-context and memory-bound. Your own number depends on context length, cache-hit rate and GPU price. Assumes roughly $2 per GPU-hour. Volume from OpenRouter's weekly leaderboard.
What is measured and what is in progress.
Measured: on Step 5's own attention layers at 32k and 100k context, on held-out code and agent-session text, the cache is 1.8× smaller than FP8 with accuracy on par with standard 4-bit cache formats. It works on live traffic as it streams in.
Built: the calibrator, with automatic quality checks and a safe FP8 fallback when a model does not pass.
In progress: full-model quality sign-off on Step 5 at 100k+ context (end-to-end accuracy and retrieval), the memory-saving GPU kernels for production serving, and the serving-engine plugin.
Serving Step 5? Become a design partner.
We measure the saving on your own traffic and GPUs before you commit.
Savings are estimates for long-context Step 5 traffic and assume roughly $2 per GPU-hour. weights.zip is not affiliated with or endorsed by StepFun or OpenRouter.
Free benchmark on your models and traffic.
- Tell us your stackThe models you serve, your context lengths and your serving engine.
- Get your numbersMemory saved, quality against your current format and capacity per GPU, measured on your prompts.
- PilotRun on one workload for 30 days. Pricing is a share of the measured saving.
Built for the models you serve.
Model names and logos are trademarks of their owners and are used only to identify the models tested. weights.zip is not affiliated with or endorsed by DeepSeek, NVIDIA, Z.ai, Meta or MiniMax.