WELVET APP 0.1.0+1 · MODEL MATRIX

What runs where.

Public snapshot for Welvet app 0.1.0+1 (v0.1 — Foundation). Fifteen models are approved in the Flutter app today. The wider catalog is tracked while GPU fusion continues.

Welvet app overview The engine

HONESTY · DONE ≠ FASTEST OR BEST

A green mark means it works.

✅ / done means that backend path exists and runs for that model in the current engineering matrix. It does not mean Welvet is the fastest implementation, the most optimised build, or the final quality bar for that device.

Round 1 is a baseline. Kernel fusion, scheduling, memory efficiency, quantisation and device-specific work continue — and the public demos will improve with every round.

SNAPSHOT · APP 0.1.0+1 · v0.1 FOUNDATION

Where things stand.

Focus right now: wiring approved models end-to-end inside the Welvet app (GUI + headless), not the later training or experiment UX.

App version0.1.0+1Approved in app15 modelsCPU done34 modelsMetal / CUDA / Vulkan triad14 modelsByteTrack Metalstill partial

Status legend.

✅ done — complete device or host path. 🟡 partial — wired but incomplete. ⬜ todo — not started. — — not applicable.

Approved Welvet models.

These are the green-list models currently approved for the Flutter app and served through the internal model host.

#	Model	Kind	CPU	Metal	CUDA	Vulkan
1	Bonsai-1.7B bonsai_1_7b	chat	✅	✅	✅	✅
2	YOLOS Tiny yolos_tiny	vision	✅	✅	✅	✅
3	ByteTrack bytetrack	vision	✅	🟡	✅	✅
4	YuNet yunet	vision	✅	✅	✅	✅
5	SFace sface	vision	✅	✅	✅	✅
6	Depth Anything V2 Small depth_anything_v2	vision	✅	✅	✅	✅
7	Google ViT Base vit_base	vision	✅	✅	✅	✅
8	RTMPose-X rtmpose_human	vision	✅	✅	✅	✅
9	DWPose dwpose	vision	✅	✅	✅	✅
10	U²-Net-P u2netp	vision	✅	✅	✅	✅
11	Bark 3.0 bark3	asr	✅	✅	✅	✅
12	Paradee-8M-v1.2 paradee	tts	✅	✅	✅	✅
13	Laya laya	other	✅	✅	✅	✅
14	D1 Omni d1_omni	other	✅	✅	✅	✅
15	MotionBricks G1 motionbricks	other	✅	✅	✅	✅

Next fuse targets.

High-level gaps ordered by closeness to the Metal / CUDA / Vulkan triad. This is a product snapshot, not a promise of ship dates.

MonoScene

Metal + CPU done; CUDA / Vulkan still todo

monoscene

Bonsai-4B

CPU + SIMD + WebGPU done; GPU fuse todo

bonsai_4b

Bonsai-8B

CPU + SIMD + WebGPU done; GPU fuse todo

bonsai_8b

Bonsai-27B

CPU + SIMD + WebGPU done; GPU fuse todo

bonsai_27b

SmolLM2-135M

WebGPU done; Metal / CUDA / Vulkan todo

smolm2_135m

Real-ESRGAN x4plus

CPU done; GPU fuse todo

realesrgan_x4plus

DeepFilterNet3

CPU done; GPU fuse todo

deepfilter_net

Silero VAD

CPU done; GPU fuse todo

silero_vad

Qwen3-ASR 0.6B

CPU + SIMD done; GPU fuse todo

qwen3_asr_06b

Qwen3-TTS 0.6B

CPU + SIMD + WebGPU fuse done; Metal / CUDA / Vulkan todo

qwen3_tts_06b

HONESTY NOTES

Backend availability, model compatibility and performance vary by device and build. A ✅ on this page means that path exists in the current engineering matrix for that model — not that every app binary ships every backend, or that every device will match a published demo timing.

Working ≠ finished. Working ≠ fastest. Working ≠ best. App 0.1.0+1 is a foundation cut; demos and fused kernels improve over successive rounds.

Back to Welvet app Demo playlist Engine overview →