Foundry is not a collection of stock models behind a graphical interface. It is a local audio production system with native Foundry model generations, custom inference, model delivery, GPU-aware orchestration, editing, mixing, DSP, and automation built to work together.
The distinction matters. A research foundation can be part of a model's history without defining the product built from it. Foundry changes the architecture, inference path, model formats, memory use, controls, and production workflow around those foundations. The result is not the same as downloading a checkpoint and placing a UI in front of it.
Foundry Speech v4 is a native speech stack
Foundry 2.0 introduced Speech v4 Large and Medium. Speech v4 advances the Qwen3-TTS foundation through extensive architectural modifications and extensions. It is built for stable long-form generation, controlled delivery, consistent speaker identity, high performance, and a low VRAM footprint.
The native speech stack covers voice generation, cloning, batching, queueing, style preparation, document narration, and the v4 zvoice format. Speech v4 Medium can run on hardware with as little as 4 GB of VRAM at some quality loss, while the Large model provides the stronger quality path on cards with more room.
Foundry has 38 delivery styles with 5 intensity levels. They cover emotion, drama, narration, and character states. A voice can be happy, angry, heroic, calm, villainous, documentary-like, exhausted, breathless, whispering, drunken, singing, and much more without creating a separate voice for each performance.
Foundry Music v4 is built for quality across consumer GPUs
Foundry 2.1 introduced the native Foundry Music v4 engine with improved sound quality, stronger long-form stability, and a much smaller VRAM footprint. It is delivered as six models: Small, Medium, Large, X-Large, Large Fast, and X-Large Fast.
Those models cover GPUs from 6 GB to 12+ GB. The highest-quality model now uses about 11 GB of VRAM instead of more than 22 GB, while the Fast variants can generate up to four times faster at lower diversity. Long tracks also benefit from spectral-flux stabilization, clearer model selection, better progress and cancellation, and more reliable automatic model choice.
Generation is not treated as a one-shot result. You can patch a weak section, blend alternate takes with spectral crossfades, separate stems, and keep refining the track without throwing away everything that already works.
The rich Speech Editor is part of the engine
Foundry 2.2 rebuilt narration around a rich document. A short sentence and an entire chapter use the same editor. Speakers, delivery styles, music, effects, custom delays, overlaps, and crosstalk can all be authored in one readable document.
Copied text keeps human-readable tags such as [Speaker tom] and [angry], so speaker and performance information remains understandable outside the editor. Complete narration no longer has to be assembled from spoken clips in Tracks, although a speaker, paragraph, or full document can still be sent there when deeper timeline editing is useful.
Preview, Play All, every export format, and Tracks actions share one sample-accurate arrangement graph. Trim, speed, pitch, volume, effects, signed delay, overlap, and DSP are calculated by the same document services. What you preview is what you hear in playback and export.
Creative AI connects ideas to production
Creative AI helps turn a rough idea into structured captions, lyrics, speech direction, and production settings. It can improve prompts, shape sections, prepare narration, and operate Foundry through its agentic and automation surfaces.
This layer is connected to the actual document, generation, arrangement, and export services. It is not a chat box that stops at writing suggestions. The same services are available through the application, API, CLI, and semantic commands.
Editing and DSP finish the work
A generated result is rarely a finished result. Foundry includes an integrated mixer, arrangement tools, patch workflows, spectral crossfades, and 32 base DSP effects. Those effects split into many presets, sub-effects, and audio modifications that can be combined into hundreds of processing options.
You can split audio into stems, repair only the part that needs work, place a voice in a room or phone call, or reshape it into something stylized. The important part is that generation, correction, layering, processing, and export remain in the same production system.
GPU support is part of model delivery
Foundry does not ask every GPU to load the same oversized model. Its model families, quantization, VRAM estimates, automatic selection, and saver modes adapt the workload to the available hardware.
- 4 GB NVIDIA: Speech v4 Medium is available with some quality loss.
- 6 to 8 GB: Speech and Music v4 run through smaller models and VRAM-saving modes where needed.
- 8 GB AMD: Full Music and Speech quality is available through the experimental Vulkan backend.
- 12 GB and above: The highest-quality Music v4 model fits alongside regular speech and production work.
- 16 GB and above: Larger Creative AI models and heavier multi-stage projects have more room.
- 24 GB and above: The largest Creative AI options, parallel work, and heavy batch production become practical.
AMD and Vulkan support remains experimental, while CUDA is still the recommended backend for NVIDIA GPUs. Foundry improves detection, VRAM reporting, verification, downloads, and recovery so the correct model and backend can be delivered without hidden setup work.
The key idea
Foundry is not a vanilla model wrapper. Speech v4, Music v4, Creative AI, rich-document narration, stem separation, editing, DSP, and automation share one local production system. The models are important, but the native engines and services that make them work together are the product.