OpenTail.Stingray 1.0.6

OpenTail.Stingray

The Unified Native Multimodal AI Engine for .NET 10
100% pure managed C# — Zero Python, Zero P/Invoke to llama.cpp, Zero external server sidecars.
Runs in-process on CPU (AVX-512/AVX2/NEON), Vulkan, or CUDA. Publishes as a single NativeAOT binary.

Built by opentail.net

Why Stingray?

In Python and C++, running modern local AI requires juggling 4–5 fragmented, heavy tools (llama.cpp for text, whisper.cpp for speech, kokoro for voice synthesis, and ComfyUI or diffusers for image/video generation).

Stingray unifies the entire local AI ecosystem into one clean, lightweight .NET library. All tensor math, SIMD kernels, KV caches, diffusion flow-schedulers, audio DSP resamplers, and tokenizers are written in high-performance managed C# that you can step into, profile, and deploy anywhere.

Superpowers at a Glance

• 💬 Text LLMs & Sparse MoE: Llama 3 / 3.1 / 3.2 / 3.3 / 4, Qwen 2.5 / Qwen 3.5 MoE, DeepSeek-V3 / R1, Gemma 3 / 4, SmolLM2. Full tool calling, JSON schema constrained decoding, and grammar masks.
• 👁️ Multimodal Vision (11+ Architectures):
  • Alibaba: Qwen2.5-VL, Qwen3-VL (3D Conv stem, M-RoPE, $2\times 2$ spatial merge).
  • DeepSeek: DeepSeek-OCR & DeepSeek-OCR2 (Dual SAM + CLIP ViT fusion, $1024\times 1024$ grid).
  • Mistral AI: Pixtral 12B (2D Continuous RoPE, SwiGLU, dynamic aspect ratios).
  • LLaVA Team: LLaVA-1.5, LLaVA-NeXT, LLaVA-OneVision (CLIP/SigLIP ViT + GELU MLP).
  • OpenGVLab: InternVL 2.5, InternVL 3, InternVL 4 (PixelShuffle $2\times 2$ downsampling).
  • OpenBMB: MiniCPM-V 2.6 (Dynamic HD 9-slice grid + 2D sinusoidal cross-attention Resampler).
  • Zhipu AI: GLM-4V, GLM-4.5V, GLM-OCR (Dual Conv2D stem, 2D M-RoPE, Conv2D patch merger).
  • NVIDIA: Nemotron-V2-VL / Nemotron-4-Nano (Learned register tokens + $2\times 2$ merge + Squared ReLU MLP).
  • Baidu / Dots: PaddleOCR-VL, Dots-OCR (2D M-RoPE + patch merger + GELU MLP).
  • Moonshot AI: Kimi K2.5 / Kimi-VL (3D learned position embeddings + 2D interleaved RoPE).
  • Google & Meta: Gemma 4 UV (gemma4uv), Gemma 4 ViT (gemma4v), Gemma 3 SigLIP (gemma3), and Llama 4 (llama4).
• 🎙️ Studio Audio & Voice Stack:
  • Speech-to-Text (ASR): Whisper (Large-v3 / Turbo), NVIDIA NeMo Parakeet FastConformer CTC, Alibaba Qwen3-ASR (0.6B/1.7B), Qwen3-ForcedAligner (word timestamps), FunASR Paraformer, SenseVoice, and Silero VAD.
  • Neural Voice & Voice Cloning (TTS): Qwen3-TTS 12Hz (with ERes2NetV2 192-dim voice cloning speaker encoder), Kokoro-82M, Chatterbox-Turbo, F5-TTS Flow-Matching DiT, CosyVoice 2.0 / 300M, Piper VITS, and MeloTTS Multilingual VITS.
  • Broadcast DSP: Rational windowed-sinc resamplers, ATSC A/85 downmixing, and TPDF dithered 16-bit/24-bit WAV exporter.
• 🎨 State-of-the-Art Diffusion & Video:
  • Image Architectures: SD 1.5, SDXL, SD 3 / 3.5 (MMDiT), FLUX.1 (schnell/dev), FLUX.2 (Klein & Kontext multi-reference), FLUX 3 (3D/4D RoPE Multimodal), and Z-Image-Turbo.
  • Acoustic Diffusion: Stable Audio 3 (Variable-Length 44.1kHz Stereo DiT).
  • Video Diffusion: Wan 2.1/2.2 Video, HunyuanVideo, and LTX-Video.
  • Real-Time Streaming: Latent Consistency Models (LCM 1–4 step) and StreamBatchPipeline (30–60 FPS real-time webcam/video streaming).
  • Exporters: Animated GIF and animated PNG sequence writers with Floyd-Steinberg dithering.
• ⚡ Zero-Latency Hardware Engine: Nanosecond SIMD evaluation (AVX-512, AVX2, ARM Neon), sub-millisecond cached GPU topology (~/.stingray/hardware_cache.json), and SmartOffloadPlanner for discrete GPUs, APUs (AMD Vega/Intel Iris), and CPUs.

Installation

dotnet add package OpenTail.Stingray --version 1.0.6

Quick Starts

1. Unified Multimodal Vision (Qwen-VL, DeepSeek-OCR, Pixtral, LLaVA, InternVL, MiniCPM, GLM-4V)

using OpenTail.Stingray.Vision;

// 1. Open any multimodal vision GGUF projector (auto-detects architecture)
using var embedder = UnifiedVisionPipeline.Open("models/mmproj-deepseek-ocr-2-q8_0.gguf");

// 2. Load and embed an image into visual token embeddings
float[] visualTokens = embedder.EmbedImageFile("document.png", out int tokenCount);
Console.WriteLine($"Generated {tokenCount} visual tokens with embedding dim {embedder.EmbeddingDim}");

2. LLM Chat with Streaming

using OpenTail.Stingray.Core;
using OpenTail.Stingray.Cpu;
using OpenTail.Stingray.Engine;

// 1. Open GGUF model and initialize hardware backend
var model = GgufModel.Open("models/Llama-3.2-3B-Instruct-Q4_K_M.gguf");
var cpu   = new CpuBackend();
using var engine = new InferenceEngine(model, cpu);

// 2. Stream tokens in real time
await foreach (var token in engine.StreamChatAsync("Explain quantisation in simple terms."))
{
    Console.Write(token);
}

3. Native TTS Synthesis & Voice Cloning

using OpenTail.Stingray.Audio;
using OpenTail.Stingray.Audio.QwenTTS;

// 1. Initialize Qwen3-TTS pipeline
using var tts = new QwenTtsPipeline();

// 2. Generate 24kHz speech with reference audio voice cloning
var result = tts.Generate(new AudioGenerationRequest
{
    Text = "OpenTail Stingray is the unified, high-performance native .NET 10 multimodal AI engine.",
    ReferenceAudioPath = "samples/speaker_reference.wav",
    OutputPath = "output.wav"
});

Console.WriteLine($"Synthesized {result.Duration.TotalSeconds:F2}s of audio to output.wav");

Verification & Provenance

Stingray rigorously validates models against real weight binary checkpoints on disk. See docs/048-model-provenance-and-real-weights-verification-plan.md for the complete provenance matrix and benchmark test runs.

License

MIT License — Copyright (c) 2026 OpenTail.