About
Back to main

Nordic Fintech Week 2026

About Henselian

Henselian is a fast way to build and test pension and insurance products, from an idea to a real number.

Everything across this booth, Design, Game and Research, runs a genuine calculation behind the scenes, not a canned demo. This exhibit is a preview for Nordic Fintech Week 2026.

How fast is it, really?

Three real models, timed for real, at three very different scales, each measured against its own real CPU baseline on the same machine. Every number below either comes from an actual measured run or is honestly labeled as a projection from one, the same way every other number in this exhibit does.

Koller 2012: a real actuarial book

A real textbook actuarial calculation (Koller, 2012): a pension and life-insurance book valued against many simulated futures. Timed for real at several sizes, then stretched out, honestly, to a full-size book: 1,000,000 policies x 10,000 futures x 30 years.

The traditional way
4.8 daysPlain code, one ordinary computer, at full size
Henselian, one processor core
10.8 daysHenselian's own language, the same computer, at full size
Henselian, graphics card
~4,181x faster than one core
3.7 minHenselian's own language, one graphics card, at full size
Seconds or sub-second, today, at demo size

Every steady-state point from 5 to 50 scenarios (512 policies, 30 years) finished in well under a second: 43ms, 169ms, 420ms. 200 scenarios -- this benchmark's own largest previously-measured point -- finished in 1.68s. Even pushed higher, to 2,000 scenarios (512 policies x 2,000 scenarios x 30 years, the model's own real unpatched native scale), it finished in 16.83s real steady-state wall-clock. This is a real, repeatedly confirmed claim across the whole 512x(5..2,000)x30 range this benchmark actually tests. It is a separate, honest claim from the full 1,000,000x10,000x30 production target below, which is NOT sub-second even in steady state -- see extrapolated.note.

Also real today, on Hensel's own compiler
900sacross 59 real GPU launches, no host-orchestration workaround
Best estimate liability36,151,610
VaR 99.5%95,224,484
Capital required (SCR)59,072,874
Total market-consistent value39,695,982

All three timed numbers are real projections, not guesses: each version of the calculation was timed at several real, smaller sizes, and the time grew in an almost perfectly straight line, so that same line was extended out to the full size above. None of the three full-size runs was actually attempted; doing the first two for real would themselves take days. The graphics-card run is different: at the smaller, real sizes actually tested above it finishes in well under a second up to a few tens of seconds, but at the full size above the honest, computed answer is still minutes, not seconds, because reading back the answer for each step is a real, mostly-fixed cost paid ten thousand times over.

Age-Integrated Society: this exhibit's own reserve

One-year SCR panel, two-member household, six integration/mortality bases, 16,384 Monte Carlo paths per basis.

CPU interpreter, all cores
1,266.88sreal, measured, same fixed scale
Graphics card
~3.0x faster
423.73sreal, measured, same fixed scale
VM-22: a 1,000,000-policy book

VM-22 Principle-Based Reserve, 1,000,000-policy book, 1,024 Monte Carlo paths per policy.

CPU interpreter, all cores
3.33 daysprojected from real smaller runs, full 1,000,000-policy book
Graphics card
~37,673x faster
7.63sreal, measured, full 1,000,000-policy book

Every benchmark on this page is a real measurement or an honestly-labeled projection from one. Nothing here is a canned demo, and nothing is invented.

Back to main