Wayner Barrios
Wayner Barrios is an individual multimodal-AI researcher and solo open-source developer, not a company. He develops models and systems connecting visual perception with higher-level reasoning, with work spanning video understanding, visual grounding, transparent reasoning evaluation, reinforcement-learning post-training, and efficient inference. He also founded Wiqonn, an AI lab focused on taking applied AI work from research into production in Latin America.
At a Glance
- Researchers and developers working on multimodal AI
- Organizations needing applied AI, software, data or infrastructure engineering
- Teams seeking local/private inference on Apple Silicon
- AI teams fine-tuning and serving models on Blackwell/DGX Spark hardware
- +1 more
AI Tools by Wayner Barrios
(1)
vllm-mlx
Local LLM Server for Apple Silicon
Discussions
No discussions yet
Be the first to start a discussion about Wayner Barrios
Latest News
Wiqonn published 'Structured Output Does Not Guarantee Visual Grounding,' an empirical study of coordinate interfaces, crop selection and answer reliability for vision-language models.
vLLM-MLX released v0.5.0; the GitHub repository identifies it as the latest release.
Wiqonn published 'Fine-Tune LLMs on a DGX Spark: LoRA + NVFP4 in Practice,' covering Blackwell GB10 fine-tuning, quantization and TensorRT-LLM serving.
MoDA was presented in the ICML 2026 program as a paper by Wayner Barrios and collaborators.
Products & Services
Open-source, native-MLX inference server for Apple Silicon that exposes OpenAI- and Anthropic-compatible APIs for language and multimodal models. It supports continuous batching, content-based multimodal prefix caching, paged cache management, vision, audio, video, tool calling and MCP workflows.
A lightweight instruction-guided channel-modulation module for fine-grained visual grounding in multimodal language models. The official guide reports under 1% additional FLOPs and implementations for LLaVA-family and Qwen3-VL workflows.
A multimodal reasoning benchmark, dataset, metrics package and process-reward training framework that evaluates whether intermediate reasoning steps are relevant, complete and correctly ordered. It contains 6,372 instances and includes Match F1 and Ordered Match F1 metrics.
An experimental end-to-end pipeline for LoRA fine-tuning, NVFP4/MXFP8 quantization, TensorRT-LLM export and OpenAI-compatible serving on NVIDIA's Blackwell GB10/DGX Spark platform.
Market Position
Wayner's work sits at the intersection of multimodal-AI research and practical open-source systems: MoDA targets visual-grounding quality, CRYSTAL targets transparent step-level reasoning evaluation, and vLLM-MLX targets native Apple-Silicon serving. In inference, vLLM-MLX is positioned as a local Apple-Silicon alternative or complement to llama.cpp, mlx-lm and hosted APIs, while retaining OpenAI/Anthropic-compatible interfaces and adding continuous batching and multimodal caching.
Leadership
Founders
Wayner Barrios
Multimodal AI researcher with a Ph.D. in Computer Science from Dartmouth College (awarded Spring 2026), advised by SouYoung Jin. His earlier research includes computer vision, video understanding and medical AI; public profiles identify him as a Research Associate at Carnegie Mellon University's Robotics Institute and as co-founder of Wiqonn.
Founding Story
Wayner Barrios founded Wiqonn to grow ambitious AI work in Latin America. His stated research-to-production vision is to connect perception and reasoning with practical, efficient systems that organizations can adopt and operate themselves.
Business Model
Revenue Model
Wayner's related venture, Wiqonn, sells custom software, data systems, AI, infrastructure, digital-transformation and team-training services. The Wiqonn site says each phase's scope, evidence plan and price are defined in writing; the named research and open-source tools are publicly available rather than presented as paid SaaS products.
Pricing Tiers
Wiqonn states that scope, evidence plan and price are defined in writing for each phase; no public tiered price list was found.
Target Markets
- Researchers and developers working on multimodal AI
- Organizations needing applied AI, software, data or infrastructure engineering
- Teams seeking local/private inference on Apple Silicon
- AI teams fine-tuning and serving models on Blackwell/DGX Spark hardware
- Latin American organizations seeking AI technology transfer and adoption support
- Research on multimodal model perception, grounding and reasoning
- Long-video moment retrieval and temporal question answering
- Evaluation of the completeness and ordering of model explanations
- Local/private LLM and multimodal inference on Apple-Silicon Macs
- Multi-user local model serving with continuous batching
- Fine-tuning and quantization of LLMs on NVIDIA DGX Spark/Blackwell hardware