BlackRiver / Local intelligence
Browser-native products, native llama.cpp releases and model-engineering research built to keep inference close to the hardware. From compact WebGPU models to 284B-class GGUF systems.
Each lab is a working product surface or a documented model release. Browser labs require WebGPU; native releases target local runtimes such as llama.cpp.
A complete private AI workspace for conversations, documents, scanned-PDF OCR, images and audio. Choose Gemma 4 E2B or E4B and run the full experience locally in your browser.
Powered by the proven BlackRiver Gemma 4 WebGPU runtime and local document pipeline.BlackRiver AI product: model management, chat, local storage, OCR, citations, multimodal attachments and resilient browser inference.
BlackRiver's daily-driver sparse MoE for coding, agents and long-horizon reasoning. In our internal real-world workflow evaluation it consistently beats the upstream Qwen3.6-35B-A3B we started from.
Base model: Qwen/Qwen3.6-35B-A3B · Training base: unsloth/Qwen3.6-35B-A3B.Created by A.I Joe and published by BlackRiver AI Ltd: Phase 10.2 curriculum post-training, exact BF16 adapter merge, MTP-preserving GGUF conversion, multimodal projector and release validation.
Two llama.cpp-ready GGUF variants engineered to preserve capability while substantially reducing refusal behavior. Flash prioritizes maximum practical context; Flash Pro uses a higher-precision IQ3_XXS foundation.
Base model: DeepSeek V4 Flash 0731 · Quantized foundations: Unsloth UD-Q2_K_XL and UD-IQ3_XXS.Created by A.I Joe and published by BlackRiver AI Ltd: refusal-direction projection, norm-preserved Q8_0 payloads, byte-range patching and full integrity audit.
Generate multilingual speech and clone a reference voice entirely in the browser. Parakeet transcribes the reference automatically before private, local WebGPU synthesis.
Core TTS: OmniVoice by k2-fsa · Automatic speech recognition: NVIDIA Parakeet · Browser ONNX packaging and community implementation: VocoLoco.BlackRiver AI implementation: unified loading flow, automatic transcription, long-form generation, interface and deployment.
Large-scale private reasoning with a dense 27-billion-parameter model compressed to 1-bit weights.
Model and low-bit research: Prism ML · Base architecture: Qwen3.6-27B · Browser foundation: WebML Community.BlackRiver AI implementation: product interface, integration and deployment.
Understand images, transcribe audio and converse with a compact multimodal model entirely in your browser.
Model: Google · Browser foundation: WebML Community and Hugging Face Transformers.js · WebGPU kernels: Fable 5.BlackRiver AI implementation: multimodal adaptation, interface and site integration.
A compact Liquid AI language model tuned for extremely fast local generation and structured tasks.
Model: Liquid AI · Browser foundation: WebML Community · Original kernel optimization credits: Fable 5 and Opus 4.8.BlackRiver AI implementation: product interface, navigation and deployment.
Generate high-quality images locally with compressed 4B diffusion models and private prompts.
Low-bit models: Prism ML · Base model: FLUX.2 Klein 4B by Black Forest Labs · Browser foundation: WebML Community.BlackRiver AI implementation: product interface, integration and deployment.
“BlackRiver AI implementation” describes the unified product interface, navigation, deployment packaging and selected workflow adaptations for third-party browser research. QwiVer and DeepRiver are separately identified as BlackRiver model-engineering releases built from fully credited upstream foundations.
QwiVer is derived from Qwen3.6-35B-A3B and retains upstream Qwen architecture credit. A.I Joe / BlackRiver AI developed the Phase 10.2 curriculum post-training, exact BF16 adapter merge, MTP-preserving GGUF release pipeline, multimodal packaging and validation. Read the complete QwiVer presentation.
DeepRiver is derived from DeepSeek V4 Flash 0731 and preserves the original Unsloth quantization credit. A.I Joe / BlackRiver AI developed the quantization-aware behavioral modification, direct Q8_0 replacement payloads, GGUF range-patching pipeline, integrity manifests and release packaging. Read the complete technical presentation.
k2-fsa created OmniVoice. Automatic reference transcription uses NVIDIA Parakeet. The browser-ready ONNX work and model packaging are credited to VocoLoco. BlackRiver AI integrated the complete local workflow, interface and deployment surface.
Prism ML created the Bonsai 27B low-bit model, derived from Qwen3.6-27B. The original browser work is credited to WebML Community.
Google created Gemma 4. The browser foundation comes from WebML Community and Hugging Face tooling; the original optimized WebGPU kernels are credited to Fable 5.
Liquid AI created LFM2.5 230M. The original browser demo is credited to WebML Community, with kernel optimization credits retained from that project.
Prism ML created the binary and ternary Bonsai Image variants from FLUX.2 Klein 4B by Black Forest Labs. The browser foundation is credited to WebML Community.
Inference happens on your GPU. Content does not need to leave the browser.
Every runtime uses real quantized weights and documented foundations, with no simulated outputs.
WebGPU makes browser AI accessible; llama.cpp brings frontier-scale GGUF models to native hardware.