Everything on this page was measured by our own team, on our own hardware. That's why every number carries an internal benchmark label — there are zero third-party reproductions yet. When external runs land, they get their own badge. Until then, treat these as our claims, not the community's.
Internal Verified — Linux x86_64 · NVIDIA CUDA (measured on an RTX 3090) Experimental — Windows WSL2 + NVIDIAWaitlist — macOS Apple Silicon (not released)
Published runs
Internal benchmark results
Two runs are public so far. Both were executed warm on an RTX 3090 (24 GB) under the Internal Verified environment class — that card is the bench machine we measure on, not a floor for running the engine; the working set below is 9.55 GB. No user data appears in this feed — capsule id, environment class, and timing only.
LTX-Video 2B · text-to-video
capsule: ltxvideo2b_t2v_768x512_49f_50step
Internal Verified
Environment
Linux x86_64 · NVIDIA RTX 3090 (24 GB)
Profile
768×512 · 49 frames · 50 steps
Success
completed · no OOM
Working set
9.55 GB resident working set · internal · RTX 3090 · not a peak that tracks the workload
Also measured
Same model at 768×448 · 49 frames · 30 steps: the same 9.55 GB over 5 runs, spread 0.00 GB. The working set does not move with resolution or step count.
Wall time
measurement in progress — render-time benchmarks are being finalized
working set measured 2026-07 · timing pendinglabel: internal · working set only
CogVideoX-2B · text-to-video
capsule: cogvideox2b_t2v_720x480_49f_50step
Internal Verified
Environment
Linux x86_64 · NVIDIA RTX 3090 (24 GB)
Profile
720×480 · 49 frames · 50 steps
Success
completed · no OOM
Wall time
measurement in progress — render-time benchmarks are being finalized
timing pending re-measurementlabel: internal
Timing is wall-clock for the render step on a warm engine (weights resident, first run excluded). Numbers are profile-conditional — resolution, frame count, step count, and model all change the result. This feed never includes prompts, file paths, account identifiers, or network addresses.
Not yet published
Re-verification in progress
These measurements exist internally but did not pass our bar for publication yet. They are listed by name only — no numbers until re-verification completes.
FP8 speedup vs fp16 baseline
held for re-measurement under the published methodology
re-verifying
Head-to-head vs other local runtimes
no comparative claim until both sides run under identical conditions
re-verifying
Known limits (Technical Preview): the engine only applies quantization where it holds up against the fp16 reference, and disables it where it does not. And to repeat the headline: there is no external reproduction data yet.
Methodology (summary)
How the published numbers were taken:
Warm runs — engine and weights resident; cold-start time is excluded and reported separately when relevant.
Fixed seed — identical seed across repeats of the same capsule.
Declared profile — resolution, frames, and steps stated on every card; no cross-profile comparisons.
The complete condition set — hardware state, driver/CUDA versions, repeat counts, and timing boundaries — will be documented in the benchmark-method document. Until it ships, this summary is the authoritative description.
Reproduce it on your GPU
The fastest way to turn "internal benchmark" into a verified number is an external run. Install free, run the same capsule, and file a compatibility report — successes and failures both count.
Free tier is enough to reproduce — no watermark, commercial use permitted. Machine-readable AI provenance is on in every tier.
Linux x86_64 + NVIDIA CUDA is the published environment (our own runs were taken on an RTX 3090). Reports from 8, 12 and 16 GB cards are the ones we most want — those are the cards the engine is built for and the ones we have no external data on. WSL2 reports are welcome as Experimental.