Built for inference clusters
Your GPU fleet, doing several times the work
TIMRUN is a drop-in replacement for vLLM or SGLang on GPUs you control. Use the same models, same hardware, and get several times the output.
Token throughput · Same models, same hardware
3.5×
Faster token throughput
At 200k tokens of context vs vLLM
2.3×
Concurrent workloads
More output from the same fleet
10×
Stateful context window
Beyond the base model limit via KV compression
Built for expensive workloads
Agent traffic is context heavy. We're built for it.
Our system compresses context mid-run in the KV cache. A prefix-only cache re-encodes everything after the cut. Our system reuses the suffix too, which reclaims data and significantly improves cache hits for throughput.
Learn more about our research →.
STANDARD RUNTIME · PREFIX CACHE ONLY
Everything after the pruned span is re-encoded from scratch.
SUBCONSCIOUS TIMRUN · PREFIX + SUFFIX CACHE
The suffix is reused too. Some of the pruned token information is preserved in the suffix.
Each strip is one agent context after a mid-conversation prune. A standard prefix cache only reuses tokens before the pruned span, so everything after it is re-encoded. TIMRUN also reuses the suffix, keeping most of the context cached.
Licensing
Terms you can put in a spreadsheet
No per-token math, no GPU-type tiers. One number times your GPU count.
Talk to Us About a Demo →