Download & Get Started
Everything you need to know before downloading S2S Portable. System requirements, prerequisites, licensing, and step-by-step setup instructions.
Everything you need to know before downloading S2S Portable. System requirements, prerequisites, licensing, and step-by-step setup instructions.
VRChat players, gamers, Discord calls, casual Zoom/Teams conversations. Free for home and personal use.
Streamers, YouTubers, podcasters who need real-time translation for multilingual content and audiences.
Call centers, international teams, professional interpreters. Commercial license required for business environments.
| Component | Requirement |
|---|---|
| OS | Windows 10 / 11 |
| GPU | NVIDIA RTX 3060 12GB+ |
| RAM | 8 GB min, 16 GB recommended |
| Storage | ~50 GB total (app + models) |
| Audio | Microphone + speakers/headphones |
| When | Needed? |
|---|---|
| Downloading S2S | Yes (~37 GB one-time) |
| Downloading NLLB | Yes (~5 GB one-time) |
| Downloading LLM files (optional) | Yes (~8.1 GB one-time) |
| Running S2S | ❌ No — 100% offline |
| Updates | Internet for download only |
Once downloaded, S2S Portable never needs an internet connection. All AI runs locally on your hardware.
© VB-Audio Software — Vincent Burel. Free donationware.
S2S Portable routes translated audio to VRChat, Discord, Zoom, and other apps through a virtual audio cable. This must be installed before running S2S.
Startup auto-test, Open Windows Sound, Quest / VRChat / Windows routing, and config.ini — separate setup guide.
Meta's NLLB translation models work seamlessly with S2S Portable out of the box. Simply drop the model files into the correct folder and S2S detects them automatically. The separate download is required solely for copyright compliance with Meta's license.
NLLB (No Language Left Behind) by Meta AI supports 200 languages. These pre-converted int8 models are optimized for speed. You only need one model to get started:
635 MB
Fastest & lightest. Best for lower-end GPUs (RTX 3050–3060).
1.36 GB
Best balance of speed and quality. Ideal for most users (RTX 3060–4070).
3.21 GB
Best translation quality. For high-end GPUs with VRAM to spare (RTX 4070+).
Extract the downloaded archive and place the model folders inside your S2S directory:
S2S Portable automatically detects and uses whatever models are present. No configuration needed.
These models are based on Meta AI's NLLB-200 and licensed under CC-BY-NC-4.0. They may only be used for personal, non-commercial purposes. For commercial translation, S2S Portable supports Qwen3 (Apache 2.0 — full commercial use permitted).
Attribution: NLLB Team et al., “No Language Left Behind: Scaling Human-Centered Machine Translation,” Meta AI, 2022. Models converted to CTranslate2 int8 format.
The link below downloads one TransferXL zip (browsers often save it as TransferXL-08vMcqLcTmRkRY.zip). Extract it to get three .gguf checkpoints, then move each file into S2S’s models folder. This is the same cutting-edge multilingual bundle power users run — without hunting files across the web. Already have another model? You can still add any multilingual LLM you want if it uses .gguf and follows the models folder rules below.
.gguf format
S2S Portable scans the models folder for .gguf checkpoints. You are not limited to our trio: place whatever multilingual LLM you prefer there, as long as the file ends in .gguf and sits directly in models (not inside nllb or any subfolder). Pick a size your GPU can load. The three files listed further down are a curated, tested starter set — strong multilingual instruction models for reasoning, rephrasing, and extra language help next to your speech translation pipeline.
No. You get the full trio so you can choose the best fit for your GPU: a featherweight option for tight VRAM, a balanced mid-size model, and a larger checkpoint when you want maximum quality. Install all files, then pick the one you want inside S2S — the others stay on disk for when you upgrade hardware or switch workloads. Your own .gguf files appear in the same list once they sit in models.
Recommended bundle — three .gguf files inside one zip (from the link below)
Qwen3.5-0.8B.q8_0.gguf — Smallest & fastest. Qwen3.5 in a tiny footprint: great when VRAM is limited or you want instant responses.Ministral-3-3B-Instruct-2512.Q6_K_H.gguf — Sweet spot. Ministral instruct tuning with excellent multilingual coverage — the workhorse most people run day to day.qwen2.5-7b-instruct-q4_k_m.gguf — Heaviest hitter. Qwen2.5 7B at solid quantization: reach for this when you have headroom and want the richest answers.Put every LLM checkpoint only in the models folder — not inside models\nllb or any other subfolder. Applies to our bundle and any other .gguf you add.
Extract TransferXL-08vMcqLcTmRkRY.zip (or whatever name your browser saved). Move each of the three .gguf files into S2S_Portable\models\ next to the app — not inside subfolders. Keep these exact filenames. Any other multilingual .gguf you download elsewhere goes in the same folder with the same rules — if S2S can load it on your GPU, it will show up in the model picker.
If S2S was open, close and reopen it once so the new checkpoints show up in the model picker.
Each weight file remains under its original open license (for example Apache 2.0 or Mistral’s terms). You are responsible for using them in line with those licenses and with your S2S license tier (personal vs commercial). When in doubt, prefer the smaller Qwen checkpoints for experimentation and keep commercial workflows aligned with your agreement.
S2S Portable is built with the following open-source technologies. We respect all licenses and give full credit to the original authors.
| Component | License | Commercial Use |
|---|---|---|
| Faster Whisper (STT) | MIT | ✅ Yes |
| Qwen3 7B (Translation) | Apache 2.0 | ✅ Yes |
| Piper TTS | MIT | ✅ Yes |
| MeloTTS | MIT | ✅ Yes |
| OpenVoice V2 (Voice Clone) | MIT | ✅ Yes |
| NLLB (Translation) | CC-BY-NC-4.0 | ❌ Non-commercial only |
| VB-Audio Virtual Cable | Freeware / Donationware | Separate install |
One-time purchase. No subscriptions. Includes 1 year of updates.
Commercial licenses use Qwen3 (Apache 2.0) as the translation engine, which is fully cleared for commercial use. NLLB models (CC-BY-NC-4.0) may not be used in commercial/enterprise environments per Meta's license terms.
Purchased a Freelancer, Business, or Enterprise license? Here's how to activate it:
Serial Key format (provided after purchase)
| Download | Size |
|---|---|
| S2S Portable v5.0 | ~37 GB |
| NLLB Models (all 3) | ~5.1 GB |
| LLM pack (3 GGUF models) | ~8.1 GB |
| Total | ~50 GB |
Typical TransferXL save names: TransferXL-08dRG4LGrXRsk.zip = S2S Portable app (~37 GB). TransferXL-08vMcqLcTmRkRY.zip = optional LLM pack (three .gguf files inside — extract first; ~8.1 GB). Your browser’s filename may differ slightly; the ID in the middle matches these official links.