ToolScore

LTX-2.5

video gen · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12

The only local model with native audio+video in one pass — voice and picture from a single 19B model, under Apache code with revenue-capped community weights (free under $10M/yr).

#4 of 14 in Video Gen

Real price
$0 · OSS
open-source
Free tier
Yes
The code is Apache; the community weights are revenue-capped — free while you stay under $10M/yr (license corr
Limits
7.0/10
$0 code (Apache) · community weights free under $10M/yr · fp8/fp4 quants
Verdict
ONLY LOCAL AUDIO+VIDEO
No official VRAM figure is published (— in our snapshot); fp8/fp4 quants are the practical path. 19B params. W

ToolScore breakdown — why 76/100

Price / Value
9.5
Token & Rate Limits
7.0
Speed / Latency
5.0
Privacy & Retention
7.5
BYOK (Own API Key)
9.5
Output Quality
10.0
Stability
7.5
Ease of Setup
8.0
Billing Transparency
7.0
Support & Community
5.0

10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →

What it does well

  • ▸Native audio+video in one pass (audio VAE confirmed) — voice and picture from a single model, unique among local rows
  • ▸Apache code with downloadable community weights and fp8/fp4 quants
  • ▸Electricity economics: re-rolls cost watts, not credits

Where it costs you

  • ▸Weights are revenue-capped: free only while you stay under $10M/yr
  • ▸The license has already been corrected once (2026-10-29) — a moving target
  • ▸No official VRAM figure published — hardware planning rides on quants

Hidden traps — the stuff the pricing page doesn't say

medium riskRevenue-capped weights

The community weights are free only while your revenue stays under $10M/yr. Cross that line and the terms change — budget for a license conversation before you budget for GPUs. The license was corrected on 2026-10-29, so re-read the current version, not a cached one.

medium riskNo official VRAM target

The VRAM cell in our snapshot is '—': fp8/fp4 quantizations are the practical path, and any hardware claim you find is community-tested, not vendor-published.

Try LTX-2.5

Know the limits before you pay: read the traps above, then decide with eyes open.

Affiliate link where available — never affects the score.

What the advertised price hides
The code is Apache; the community weights are revenue-capped — free while you stay under $10M/yr (license corrected 2026-10-29). Cost of running is electricity, ≈$0.03/hour for a 170 W card at the US average $0.18/kWh.
Sources & freshness

Data as of 2026-11-12. Prices and limits change monthly — report anything stale. How we verify: methodology.

Closest alternatives

LTX-2.5 review

In-depth analysis · last updated 2026-11-12

One pass, two senses

LTX-2.5 is the only local row with native audio+video in one pass — the audio VAE is confirmed, and voice and picture come out of a single 19B model. Every other open model in this comparison generates silent picture; pairing them with voice means a separate TTS step and a sync pass. Here that whole layer collapses into the render. For local pipelines that iterate, that is not a convenience feature — it is an entire workflow removed.

The license asterisk

The code is Apache; the weights are community and revenue-capped: free while you stay under $10M/yr, with the license corrected as recently as 2026-10-29. That correction cuts both ways — the terms are public and dated, but a license that moves once can move again. Read it before betting a pipeline on it, and re-read it before you cross the revenue line.

Running it

No official VRAM figure is published — our snapshot's cell stays '—'. The practical path is fp8/fp4 quantizations, community-tested rather than vendor-blessed. The economics once it runs are the local standard: a 170 W card draws ≈$0.03/hour at the US average rate, and a rejected take costs a re-roll, not a credit.

How the pillars land

Price/Value 9.5 — audio+video for watts is the strongest per-feature value in the local table. Length & resolution caps 7 — no published hardware or caps figure; quants only. Render speed 5 — neutral: no render-time figure appears in our snapshot. Training use & license 7.5 — Apache code, but the community weights carry a $10M/yr revenue cap. API access 9.5 — the weights and quants are downloadable; there is no key because there is no gate (yet). Feature set 10 — the only local audio+video single pass. Track record 7.5 — a license already corrected once is a young, moving target. Ease of use 8 — the quant path exists, but no official VRAM target means you plan by experiment. Pricing transparency 7 — the license terms are public and dated; the hardware envelope is not published. Support & docs 5 — neutral: no support-channel facts in our snapshot.

Verdict — 76/100

Second place locally, and the only local model that speaks. If your pipeline needs voice with picture and you want it on your own GPU, this is the row — just carry the $10M license cap and the unpinned hardware figure as known unknowns.

From the insights