Research, product updates, and engineering insights.
September 2026
Remembering a fact does not mean knowing who it belongs to. A retention experiment and the questions operators should ask before serving reused state.
Read more → ProductSeptember 2026
We tested governed semantic KV reuse for open weight models on customer controlled GPUs using a research workload with more than 100 entities and 14.3 million recorded token events.
Read more → ProductJuly 2026
A single-page calculator that models your fleet as three outcomes: semantic reuse hits, exact-prefix hits, and misses. It charts cumulative cost avoided over twelve months, and the gap between the lines is what semantic reuse adds on top of the exact-prefix caching your engine already does.
Read more → ProductJune 2026
Billions are flowing into world models, so it is fair to ask whether semantic cache reuse makes them cheaper to run. The honest answer is: only in specific cases. JEPA-style latent predictors have no token KV cache to reuse; autoregressive world models and their planning loops do. Here is where the connection is real, where it is a lazy analogy, and why the hard part was never adding a cache.
Read more → ProductJune 2026
Exact prefix caching only reuses work when requests start with the same bytes. Change the system prompt, the RAG template, or the section order, and the GPU recomputes from scratch. A two-minute look at how Synapse routes long-context traffic toward reusable KV, cuts cold prefill, and turns missed reuse into fleet capacity.
Read more → ProductJune 2026
Exact prefix caching misses repeated work hidden behind different prompt framing. Synapse routes long-context traffic toward reusable KV to reduce cold prefill, improve TTFT, and raise GPU efficiency.
Read more →See how semantic context routing can reduce cold prefill, improve TTFT, and raise GPU efficiency in your own inference fleet.
Request Demo