Every CS2 grenade falls at exactly 320 units per second squared — 0.4× the 800 u/s² players fall at — and grenades have zero air drag. The same constant governs smokes, HE, flashbangs, decoys, molotovs and incendiaries: there is no per-type gravity in CS2. It was measured from 666 flight segments across four maps and then proved by replaying those flights through a float32 simulation, whose worst position error against the server was 0.0298 units.
Derived from grenade trajectories in 2026 tournament demos across four maps, then confirmed tick-exact against controlled captures on CS2 build 1.41.6.8/10847.
Plays in your browser · no download · no account
A grenade in a demo is a stream of positions, one per tick, each rounded to the nearest 1/32 of a unit. Differencing that stream twice to get acceleration amplifies the rounding instead of the signal: the second difference of a 1/32 u grid at 64 Hz quantizes in steps of (1/32)/dt² = 128 u/s², so a naive measurement of a 320 u/s² constant reports ±57 u/s² of noise and no convergence at all.
What works is fitting whole ballistic segments. A least-squares parabola over each uninterrupted flight averages the quantization away instead of amplifying it, and the per-segment fits collapse onto a single value with a median absolute deviation of 0.13. Only pre-first-bounce flight is used, because anything after a contact may be rolling rather than flying.
A converged fit is a derivation, not a proof. The proof is replay: the whole corpus is run back through a float32 simulation with only genuinely unknown quantities left free, and the requirement is that residuals stay at the network quantization level and do not grow with segment length. A residual that grows with length means the model is wrong no matter how small its mean is.
The integration ORDER is a separate fact, and it could not be derived from demos at all — a position corpus absorbs the substep order into the initial velocity, so every candidate order fits equally well. It took controlled captures on the current build to resolve: two 1/128 s substeps per 64 Hz tick, gravity subtracting 2.5 u/s at each, and z advancing by the average of the pre- and post-gravity velocity. Only that order replays the recorded float32 streams bit-identically.
The flight, the collisions, the bounces and the moment a smoke comes to rest all replay tick-exact on the current build. The delay between coming to rest and the volume appearing is not part of that gate.
Other grenade types share the flight law — that is measured — but their detonations, molotov slope behaviour at capture precision, and entity contacts are outside the controlled matrix. No 1:1 claim is made for them.
Last updated 2026-07-21 — the date of the measurement, not of this page's rendering. What the confidence tiers mean. Memorin is not affiliated with or endorsed by Valve; Counter-Strike 2 is a trademark of Valve Corporation.