Refresh Vontra organization profile

README.md

@@ -6,23 +6,44 @@ pinned: false
 
 # Vontra
 
-Solo AI tinkerer publishing practical local-inference artifacts: MLX quants, serving notes, and the small bits of glue that make models easier to run on real machines.
+### Local AI, real hardware, and useful MLX releases for the community.
 
-No startup theatre. No mystery benchmark fog. Just useful experiments, clean model cards, and attribution back to the original authors.
+Hi — I’m one person with a strong interest in local AI.
 
-## Current Shelf
+I spend my time trying models on hardware I actually own, working through the awkward compatibility problems, and sharing the useful results. Most of my work here is focused on **MLX and Apple silicon**: model conversions, quantized checkpoints, architecture support, serving notes, and the occasional bit of glue needed to make a model run properly.
 
-| Model | Format | Best for |
-|---|---:|---|
-| [Laguna-S-2.1-MLX-8bit](https://huggingface.co/Vontra/Laguna-S-2.1-MLX-8bit) | MLX 8-bit | Highest-fidelity local run |
-| [Laguna-S-2.1-MLX-6bit](https://huggingface.co/Vontra/Laguna-S-2.1-MLX-6bit) | MLX 6-bit | Middle ground between size and quality |
-| [Laguna-S-2.1-MLX-4bit](https://huggingface.co/Vontra/Laguna-S-2.1-MLX-4bit) | MLX 4-bit | Smaller local footprint |
+There is no big company or team behind this page. I’m simply exploring what local hardware can do and trying to make that work useful to other people.
 
-## What Gets Published
+## The hardware
+
+- Mac Studio
+- MacBook Pro
+- NVIDIA DGX Spark
+
+The Apple machines are where most of the MLX work happens. The DGX Spark gives me another local platform for comparison and experimentation.
+
+## What I share
 
 - MLX conversions and quantized checkpoints
-- Apple Silicon serving notes
-- OpenAI-compatible local server recipes
-- Honest caveats from running models on actual hardware
+- Support for models that do not yet work cleanly with stock MLX-LM
+- Practical oMLX and local-serving notes
+- Measured memory and generation results when available
+- Clear attribution to the original model creators
+- Honest caveats when something is experimental
+
+## Current MLX releases
+
+| Model family | Available builds |
+| --- | --- |
+| **DeepSeek V4 Flash** | [MXFP4 MLX](https://huggingface.co/Vontra/DeepSeek-V4-Flash-0731-MXFP4-MLX) |
+| **Ling 3.0 Flash** | [4-bit](https://huggingface.co/Vontra/Ling-3.0-flash-MLX-4bit) · [8-bit](https://huggingface.co/Vontra/Ling-3.0-flash-MLX-8bit) |
+| **Laguna S 2.1** | [4-bit](https://huggingface.co/Vontra/Laguna-S-2.1-MLX-4bit) · [6-bit](https://huggingface.co/Vontra/Laguna-S-2.1-MLX-6bit) · [8-bit](https://huggingface.co/Vontra/Laguna-S-2.1-MLX-8bit) |
+| **Solar Open 2 250B** | [4-bit](https://huggingface.co/Vontra/Solar-Open2-250B-MLX-4bit) · [6-bit](https://huggingface.co/Vontra/Solar-Open2-250B-MLX-6bit) · [8-bit](https://huggingface.co/Vontra/Solar-Open2-250B-MLX-8bit) |
+
+## Why I publish this work
+
+Local AI gets better when people share what worked, what failed, and what had to be changed along the way. If one of these releases saves someone else a long conversion, a difficult compatibility fix, or a few hours of testing, then publishing it was worthwhile.
+
+All base models belong to their original authors and are released under their respective licences. Please check each model card for upstream attribution, usage instructions, trust requirements, and limitations.
 
-Everything is shared best-effort, with clear credit to the original model authors where applicable.
+If you find a problem, open a discussion on the relevant model repository. Useful reports and reproducible results are always welcome.