OpenStems/OpenStems-Models

StemDeck Ultra model resources

Updated 2026-11-05. This repository contains the complete current model-weight and model-constant set used by StemDeck Ultra for Windows: 17 files, 3,705,369,284 bytes. OpenStems is the project/repository name; the application is StemDeck Ultra.

ASR Turbo is uploaded here: ggml-large-v3-turbo.bin — 1,624,555,275 bytes, original upstream F16 GGML artifact. Download ASR Turbo.

Browse all current files · Exact manifest · Version-folder documentation

Current application mapping

Application choice	Model/checkpoint	Stored format	CPU / NVIDIA GPU
Fast	B400, selected step 38200	GGUF	Same weight file and SHA256
Pro (default)	B62, selected step 27000	GGUF	Same weight file and SHA256
Ensemble	Karaoke followed by SW	Two GGUF files	Same two weight files and SHA256
Lyrics ASR	Whisper large-v3-turbo only	Original F16 GGML .bin	Same ASR weight file and SHA256
Optional 48kHz enhancement	LavaSR v2	Original FP32 .pt asset	Matching native enhancement integration
Beat analysis	Beat-this final0	.ckpt	Original application analysis asset
Song-section analysis	All-In-One Harmonix, eight folds	Eight .pth files	Original application analysis assets

CPU/GPU selection changes the matching native execution backend and configuration. It does not require a second set of separation or ASR weights. GGUF is a container format and does not itself imply integer quantization. This release adds no INT8 models.

Separation and ASR are integrated through native inference libraries. The .gguf files require the matching audio-model engine and manifests; the extension alone does not establish compatibility with an unrelated GGUF runtime. Analysis assets retain the original application analysis path.

The 48kHz option is independent of Fast/Pro/Ensemble and is off by default. The application follows its export sample-rate settings.

Complete downloadable inventory

All links below point to the immutable model-content revision f2da66683f7a0fddcd772ccd5fbf0b460378daef. File SHA256 values and relative installation destinations are recorded in models.json.

Purpose	File	Bytes	SHA256
Fast separation	B400-step38200.gguf	206,407,584	293af89efaf9650be656eeac3567ecf1368effae9b9e8660dda0561c44c68690
Pro separation	B62-step27000.gguf	812,481,984	1ded8095f4f99602766d8f60cf692eebd48da6bb9e8bcafb92b2c3b17468fbc3
Ensemble first stage	karaoke.gguf	204,258,848	4804181ac76c3d4fc9b282884cb7b6785dd517ea5e547ee90a4d9bb5eccea338
Ensemble second stage	sw.gguf	698,768,832	68bed29111b6414a5b441bd147b341dbf311a5bbaf7eac1c1073a6a82e21eb72
ASR / lyrics	ggml-large-v3-turbo.bin	1,624,555,275	1fc70f774d38eb169993ac391eea357ef47c88757ef72ee5943879b7e8e2bc69
Optional 48kHz enhancement	lavasr_v2.pt	56,389,364	8ca5a4c9114761419de20a1d8231337c3ec77e5759a206b3cfa57fbe401e0911
Beat analysis	beat_this-final0.ckpt	81,058,141	8c328b45f59d8dd3dff219253ff6a8d6482be57d0133a29140e2febbf8eb8331
B62 analysis constants	b62-hamming.safetensors	1,411,288	20e08e4596bd9e4b07245b749b58410821a43fb037d232d7e07c87ac4e2d0da0
Karaoke / SW constants	bs_constants.safetensors	8,833,400	15380c3bb19d15872e4934232079bd6c07a4bdf828bbc40d8ce3a9bc08fb5c4d
Song-section analysis fold 0	harmonix-fold0-0vra4ys2.pth	1,400,571	0db596dfb0995f41d62f6267d76a9d54c046f1649bd35e1dbeca0c5f9a7b8acd
Song-section analysis fold 1	harmonix-fold1-3ozjhtsj.pth	1,400,571	607b4864eaabccd4f6951410d5ea6676666ad1cfa6ce2e18bc99f3824b77c3b7
Song-section analysis fold 2	harmonix-fold2-gmgo0nsy.pth	1,400,571	67c28ec7ca612d365e5fcb152f6530eb53e34c4303940afb30fd17a295c958a4
Song-section analysis fold 3	harmonix-fold3-i92b7m8p.pth	1,400,571	446d05672b2669cdb7d9b8bbbc228a4f3592484335a7995cca89a744e3950ddc
Song-section analysis fold 4	harmonix-fold4-1bql5qo0.pth	1,400,571	96327222ab9b33c005388b788ba602f529aa03cc4dde1c859c011f1004490493
Song-section analysis fold 5	harmonix-fold5-x4z5zeef.pth	1,400,571	c87f162a1090290bcbb863688746784919cf912c56744fb3d81a1a7dba3b114e
Song-section analysis fold 6	harmonix-fold6-x7t226rq.pth	1,400,571	13e7e639bf30fcd3fa049ad863088773a4e57b06d7c68581385b3ef625c60688
Song-section analysis fold 7	harmonix-fold7-qwwskhg6.pth	1,400,571	b888b7fba95efa3d5a37ca013fe7efe2741621ea9ca335e6b2723e4fd0fade57

ASR identity and provenance

The ASR file is not a GGUF conversion and not an INT8 variant. It is the unchanged upstream F16 large-v3-turbo GGML asset mirrored for the application's native ASR integration.

• File: ggml-large-v3-turbo.bin
• Bytes: 1,624,555,275
• SHA256: 1fc70f774d38eb169993ac391eea357ef47c88757ef72ee5943879b7e8e2bc69
• Upstream artifact provenance: ggerganov/whisper.cpp, fixed source revision.

The application uses this one Turbo model for CPU/GPU lyrics recognition. No additional ASR model or separate alignment model is included in this release.

Download and integrity

The application installer downloads the required resources and checks the complete file identities. Integrations should pin f2da66683f7a0fddcd772ccd5fbf0b460378daef, use the manifest, and verify SHA256 before passing files to the matching native engine.

For example, the official Hugging Face CLI can download the ASR artifact directly:

hf download OpenStems/OpenStems-Models ultra-20261105/assets/1fc70f774d38eb169993ac391eea357ef47c88757ef72ee5943879b7e8e2bc69/ggml-large-v3-turbo.bin --revision f2da66683f7a0fddcd772ccd5fbf0b460378daef

Checkpoint step counts identify the selected saved separation checkpoints. This resource card does not claim a new quality score or a hardware-independent speedup. Hardware-specific performance and numerical validation belong to the application benchmark records.

Historical resources

Older root-level T400/B400/Karaoke/SW ZIPs and model-files.json are retained for reproducibility of previous application releases. T400 is not the current Fast choice, and the old Fast/Pro/Ultra mapping does not describe StemDeck Ultra. Current Fast/Pro/Ensemble and all ASR/analysis resources are listed above.

Licensing and repository scope

Original model and component terms continue to apply. Format conversion, mirroring, and packaging do not relicense the weights. Consult the corresponding upstream model terms and the application's component notices. A single blanket license is not assigned to this mixed resource collection.

This repository contains model weights, model constants, and their documentation. It does not contain user recordings, training datasets, private inference SDK binaries, machine configuration, or authentication material.

中文说明

本仓库已包含 StemDeck Ultra 当前使用的全部 17 项模型权重及常量，其中包括 ASR 的原版 F16 large-v3-turbo。ASR 文件和直接下载入口已在本页顶部列出，不只在版本目录中说明。

当前档位：Fast = B400 38200 步；Pro = B62 27000 步，默认 Pro；Ensemble = Karaoke → SW。分离模型的 CPU/GPU 共用同一份 GGUF 权重；ASR 的 CPU/GPU 共用同一份 F16 GGML .bin。没有新增 INT8 模型或独立对齐模型。

LavaSR 48kHz 为独立可选开关，默认关闭。节拍和歌曲结构分析所需的 Beat-this、八个 All-In-One fold，以及两份模型常量，也都已上传并列入清单。旧根目录 ZIP 为历史资源，当前版本请使用本页固定版本链接和 models.json 校验。

StemDeck Ultra Mac V8 model assets

Mac-native converted F32 Fast/Pro and frozen Karaoke/SW model-only archives are documented in ultra-mac-20261105-v8-r5/models.json. These preserve model precision and do not contain the private native engine SDK. Existing Windows and original model paths are unchanged. Original licenses continue to apply.