Lightricks Releases LTX-2.5, a 22B-Parameter Open-Weight Video and Audio World Model
Lightricks has released LTX-2.5, an open-weight world model that generates synchronized video and audio from text, image, and video inputs. The 22B-parameter model adds native multishot generation, a new diffusion video decoder, and a custom Gemma4 12B text encoder.
Lightricks Ships LTX-2.5 With Multishot Video Generation
Lightricks has released LTX-2.5, an open-weight model for generating synchronized video and audio from text, image, and video inputs. The model is built around a 22-billion-parameter diffusion transformer (DiT) and is designed for local execution and fine-tuning rather than exclusive API access.
The headline addition is native multishot generation. Previous LTX versions produced a single continuous shot; LTX-2.5 can generate connected scenes in one pass, holding character identity, environment, lighting, voice, and visual style across cuts, according to Lightricks.
What Changed
Lightricks lists several architectural updates alongside multishot support:
- Diffusion fidelity rendering: compute is allocated dynamically by scene complexity rather than a fixed compression rate per scene, per the company.
- New diffusion video decoder: replaces the previous VAE reconstruction stage, which Lightricks claims produces sharper faces, textures, and on-screen text along with fewer motion artifacts.
- Custom Gemma4 12B text encoder: intended to preserve detail in complex, longer prompts involving multiple characters, camera moves, and actions.
- Prompt enhancer: expands short prompts into more detailed cinematic instructions at low additional compute cost.
- Duration predictor: an optional module that infers clip length directly from the prompt instead of requiring a fixed frame count.
- Improved distilled checkpoint: Lightricks claims the distilled model retains substantially more of the full model's visual quality, prompt adherence, and motion consistency than prior distilled versions.
Distribution and Licensing
LTX-2.5 ships as a split, ComfyUI-aligned pack — separate .safetensors files for the transformer, text encoder, VAE, and duration head — rather than a single monolithic checkpoint. Variants include bf16, Comfy int8 (convrot), and NVFP4 quantized transformers, the latter supported on Blackwell GPUs via ltx-kernels.
The model is distributed under the LTX-2.x Community License. Companies with under $10 million in annual revenue can use it commercially at no cost; fine-tune transfers may require a paid license under the same terms. Organizations above that revenue threshold must negotiate a separate Paid Commercial Use Agreement covering full weights, engineering support, LoRAs, and deployment flexibility. Pricing for that tier is not disclosed. Revenue is measured across the full entity, including subsidiaries and affiliates under common control.
Installation runs through Lightricks' ltx-pipelines Python package (Python 3.12+, CUDA 12.7+, PyTorch ~2.7), with weights downloadable via the Hugging Face CLI. The distilled pipeline still depends on a spatial upscaler hosted in the separate LTX-2.3 repository, which is not yet bundled into LTX-2.5.
No formal benchmark scores were published alongside the release; all quality claims (sharper detail, better motion consistency, improved distillation fidelity) come from Lightricks and have not been independently verified.
What This Means
LTX-2.5 pushes further into the open-weight video generation space Lightricks has occupied since earlier LTX releases, competing against closed systems like Runway Gen and Google's Veo by offering self-hosted deployment and a revenue-tiered free commercial license. Multishot generation and audio-video synchronization in one pass are meaningful production features — scene-to-scene consistency has been a persistent weak point for open video models. The split-checkpoint, ComfyUI-first packaging signals Lightricks is optimizing for the existing creator-tooling ecosystem rather than a single unified API product. Whether the distilled model's claimed quality retention holds up under independent testing will determine if LTX-2.5 becomes a practical alternative to heavier full checkpoints for real-time or lower-VRAM workflows.
Related Articles
Paradigm Intelligence launches Phai-A1 language model and Phai-M1 video model, trained on Huawei Ascend
Paradigm Intelligence (范式智能, formerly 第四范式) has unveiled two in-house models: Phai-A1, a language model that runs on a single RTX 3090, and Phai-M1, a video model with synchronized audio. The company says both were trained on Huawei Ascend hardware, with open weights coming soon.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Cloudflare releases open-weight Clairic 27B and new Clairic-flash 9B for data extraction
Cloudflare announced on November 18, 2026 an updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash. Both have open weights on Hugging Face and are available as hosted Workers AI endpoints.
RadixArk releases DaisyRose-27B, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on 1,200 curated agent tasks. According to RadixArk, it raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than the base model on all seven reported benchmarks. Weights ship under Apache 2.0 in BF16, FP8 and NVFP4.
Comments
Loading...