Tempolor v4.5 is Tempolor's current flagship music-generation model. Architecturally, building on previous model versions, the bottleneck has been redeveloped; in terms of the generation paradigm it stays consistent with real music creation, producing 48kHz stereo high-quality music through a hierarchical, progressive architecture. This version decomposes music generation into learning and generation tasks at different levels, achieving coarse-to-fine structured generation: the high level handles musicality and structural organization, the middle level handles semantics and content expression, and the low level handles acoustic detail and high-fidelity reconstruction. It represents the current mainstream paradigm — a hierarchical path 'from high-level semantics to fine-grained acoustics' — balancing musicality and audio quality.

In terms of controllability, beyond ordinary generation, the model further supports precise Remix rewriting, fine-grained audio editing and more.

Tempolor v4.5 is a continued iteration of the fourth-generation hierarchical, progressive architecture, with substantial improvements to both basic music generation and Cover generation. This version shows clearer layering and spatial separation across drums, bass, harmony and vocals; on slow-tempo, relaxed styles its emotional expression and arrangement texture are especially refined. When you adjust instrument timbre, harmony or vocals via tags or prompts, the model responds closer to intent, making the process more predictable and the result more likely to land in one pass.

Basic music generation sees a major upgrade in the expression of musicality and musical semantics

• A stronger musical foundation: Further optimized model architecture and audio representations, with markedly more precise training-data annotation, giving the model a more accurate understanding of music.

• Steadier, more accurate generation: A further breakthrough in post-training: controllable responses are more accurate, output is more consistent, and the dynamics feel closer to real music.

• Stronger control and re-creation: Our self-supervised music representation model (SSL Model) was redesigned to balance the disentangling and fusion of timbre, melody and rhythm — laying the groundwork for precise Remix and song Cover.

A major upgrade to song Cover: cleaner timbre, stronger instruction following, and more stable melody reconstruction.

• Precise Remix rewriting: Reshape existing tracks at the level of style and arrangement, letting you explore many directions for the same material in no time.

• Returning to the motif: Even with difficult material — cover lyrics that differ in length from the original audio, or irregular song sections — the model converges back to the main melody quickly, avoiding loose melodic lines and leaving more dependable room for re-creation.

Based on Chinese and English test sets (30 each, 60 total), compared against Mureka v9, Suno v5.5 and MiniMax V2.6, covering the Meta Audiobox Aesthetics and SongEval evaluation systems.