Kandinsky 6.0 vs SkyReels V4 vs LTX 2.5: Best Free Open-Source AI Video Model with Audio in 2026
Open-source AI video just had its biggest week of 2026. Kandinsky Lab released Kandinsky 6.0 on October 6 — an MIT-licensed model that generates synchronized video and audio in one pass, with a 3B Lite variant that runs on a consumer RTX 4090. For creators who want video+audio without API fees or proprietary licences, the choice just got more interesting.
Here's how the three leading free, open-source video models compare — and which one to use for each job.
---
The Short Answer
| Model | Audio | Consumer GPU | Licence | Best For |
|---|---|---|---|---|
| Kandinsky 6.0 Lite | Yes — 44 kHz sync | ✅ RTX 4090 | MIT | Local inference, commercial use, budget creators |
| Kandinsky 6.0 Pro | Yes — 44 kHz sync | ❌ Cloud API | MIT | Maximum quality audio+video, production work |
| SkyReels V4 | Yes — #1 T2V-audio | ❌ High-VRAM | Permissive | Highest audio sync fidelity, cloud-first |
| LTX 2.5 | ❌ Silent video | ✅ Consumer GPU | Custom | Fast visual-only clips, highest throughput |
---
Kandinsky 6.0: The New Entry Point for Audio+Video
Released: October 6, 2026 · Licence: MIT · From: Kandinsky Lab
Kandinsky 6.0 is the first open-source video model to combine synchronized audio generation with consumer GPU compatibility. The Lite (3B) variant fits on an RTX 4090, bringing audio+video generation within reach of individual creators for the first time — without cloud bills or API rate limits.
Key specs
What Kandinsky 6.0 does well
The model excels at scene-audio lock: audio sources are placed spatially and arrive at the correct time relative to visual events. A hammer strike rings when the hammer hits. A door slam sounds precisely at the moment of contact. This is not background music laid over a clip — the audio model understands the causal relationship between visual action and sound.The Lite variant sacrifices some fine-grained audio complexity (multi-source mixing is slightly less precise than Pro) but handles single-source audio — one musical instrument, one environmental sound, one impact — at surprisingly high quality.
Sample prompts for Kandinsky 6.0
Nature — Thunder Strike: > A static wide-angle shot of an open wheat field during a storm: a lightning bolt strikes a lone oak tree at the frame's left edge. Audio: the flash is silent, then 2.1 seconds later the full thunderclap arrives — a sharp crack followed by a 4-second rolling echo off distant hills. Rain continuous underneath. Physically accurate audio delay. 5 seconds.
Music — Classical Guitar: > Close-up of a classical guitarist's right hand over the sound hole, fingers plucking an ascending arpeggio. Each note visually triggered as a finger releases the string. Audio: the guitar's warm cedar body resonance fills a small room, each note's attack and decay precisely timed to the visual finger lift, a faint room reverb of 0.6 seconds. 5 seconds.
Urban — Street Rain: > A low street-level shot of a pedestrian crossing in Tokyo during a downpour: boots splash through a puddle at the 2-second mark. Audio: the continuous percussion of rain on pavement, the specific splash of the boot, traffic hiss on wet asphalt in the background. The splash is locked to the visual impact. 5 seconds.
---
SkyReels V4: Still the Benchmark for Audio Sync Quality
Released: April 2026 · Licence: Permissive open-source · From: SkyWork AI
SkyReels V4 arrived as the first open-source model to ship synchronized audio in April 2026 and remains the benchmark for audio sync fidelity six months later. Its 30-parameter architecture (larger than Kandinsky 6.0 Pro) produces more precise synchronization in complex multi-source scenes — a crowd scene with 20 distinct audio sources, a construction site with overlapping machinery, a street market with spatially correct vendor calls.
The tradeoff: SkyReels V4 requires enterprise-class GPU infrastructure for local inference, making it primarily a cloud API model for most creators.
Where SkyReels V4 still wins
Sample prompts for SkyReels V4
Documentary — Market Scene: > A walking-pace shot through a Marrakech souk: spice stalls, a vendor pouring tea, a metalworker hammering copper, children running past. Audio: a full 5-layer soundscape — tea pour at close range, metalwork ring with delay, Arabic conversation fragments, a donkey bell 30 metres away, the call to prayer beginning at second 8. Each source is spatially locked to its visual position. 15 seconds.
Nature — Glacier Calving: > A boat-mounted camera captures a 40-metre section of glacier face fracturing and falling into the sea. Audio: 2 seconds of silence after the visual fall (the acoustic delay over 600 metres of open water), then the sharp crack arrives, followed by the sustained rumble and the wave hitting the hull. Physically accurate timing. BBC Earth quality. 12 seconds.
---
LTX 2.5: The Speed Champion for Silent Video
Released: May 2026 · Licence: Custom (non-commercial research) · From: Lightricks
LTX 2.5 doesn't generate audio — but it's the fastest high-quality open-source video model available. The 22B architecture generates clips at 3× the throughput of SkyReels V4 on equivalent hardware, making it the right choice when you need volume (social media A/B testing, storyboarding, visual reference) and don't need synchronized sound.
It also runs on consumer GPUs — an RTX 4090 handles it at quality levels approaching cloud-inference alternatives.
When to choose LTX 2.5
Sample prompts for LTX 2.5
Fashion — Fabric in Wind: > A slow-motion close-up of a silk dress in deep burgundy, fabric catching a sustained wind from the left: each fold and ripple of the fabric is physically accurate, the material's weight and drape visible in every movement. Golden late-afternoon light catches the sheen. No audio needed — pure visual. 5 seconds.
Architecture — Sunrise on Glass Tower: > A static camera captures the first 5 seconds of sunrise light hitting a modern glass tower: the reflection moves from bottom to top of the building as the sun rises, each floor reflecting a slightly different angle of orange-gold light. Time-lapse compressed to 5 seconds. No audio. Architectural photography quality.
---
The Decision Framework
Use Kandinsky 6.0 Lite if:
Use Kandinsky 6.0 Pro (via API) if:
Use SkyReels V4 if:
Use LTX 2.5 if:
---
Open-Source Video in October 2026: The State of Play
Six months ago, the open-source video category had one audio option (SkyReels V4) and a growing set of high-quality silent models. Kandinsky 6.0 changes the access equation: synchronized audio is now available on consumer hardware for the first time.
The remaining gap is clip length. All open-source audio+video models top out at 5–15 seconds. For longer-form synchronized content, proprietary APIs (Kling 4.0 at 30 seconds, Veo 4) still hold the advantage. That's the open-source frontier for 2026's remaining months.
---