Released on August 22, 2026, Trama v1.0.7 is a free offline stem-separation software for Windows, macOS, and Linux that isolates vocals, drums, bass, and other audio tracks directly on local hardware without cloud processing, utilizing Meta’s htdemucs_ft model to complete a three-minute track in one to two minutes via an NVIDIA GPU.
The utility of local audio separation has shifted dramatically as open-source models outpace commercial web APIs in both privacy and raw execution speed. Engineers no longer need to route sensitive multitrack stems through third-party servers. Trama v1.0.7 answers this demand by packaging Meta’s htdemucs_ft architecture into a native desktop binary.
Under the hood, performance scales aggressively based on hardware architecture. When running calculations through an NVIDIA graphics processing unit, processing a standard three-minute song in High mode takes roughly one to two minutes. Swap that workload to a CPU, however, and the timeline stretches to a sluggish 10 to 15 minutes.
Hardware support remains sharply bifurcated across operating system boundaries. AMD graphics accelerators currently lack native hardware acceleration support within the application framework. Similarly, systems powered by Apple Silicon route processing tasks entirely through the CPU rather than tapping into the unified memory’s neural engine cores.
Workflow Speeds and Ecosystem Footprint
Batch processing capabilities allow audio engineers to queue multiple files simultaneously. Trama accepts WAV, AIFF, MP3, and FLAC inputs, exporting the resulting isolated stems exclusively as 32-bit Float WAV files at a 44.1 kHz sample rate. That ensures maximum headroom for downstream mixing in a Digital Audio Workstation.
Installer sizes vary wildly depending on the target platform. The Windows 10/11 x64 package demands a 3.1 GB download. Linux x86_64 users pull down 3.9 GB. Meanwhile, the macOS version built for Apple Silicon weighs in at a much leaner 500 MB.
The Evolving Landscape of Offline Audio Tools
This release drops straight into a fiercely competitive field for local audio manipulation. On August 21, 2026, details emerged regarding StemDeck, an open-source platform licensed under Apache-2.0. Operating currently in an alpha state, StemDeck leverages the htdemucs_6s model to split audio into six distinct stems, including dedicated tracks for guitars and pianos. StemDeck also accepts YouTube URLs directly and automatically hooks into CUDA for NVIDIA cards or Metal Performance Shaders (MPS) for Apple Silicon.
Just a day prior, on August 20, 2026, zplane debuted PEEL STEMS 2. That commercial software targets DAW integration head-on, boasting a 49 percent reduction in processing latency to achieve real-time stem separation. Unlike subscription-heavy audio tools, PEEL STEMS 2 utilizes a perpetual license model, though latency penalties still hit Intel-based Mac hardware harder than Apple Silicon or Windows rigs.
As local inference models grow more efficient, tools like Trama demonstrate that high-quality stem separation no longer requires cloud infrastructure or recurring SaaS fees. Engineers retain complete data sovereignty while keeping complex multitrack workflows entirely on their own iron.
