Embed logo + quickstart + tier matrix

README.md

@@ -1,104 +1,69 @@
 ---
-title: Outlier
+thumbnail: https://huggingface.co/Outlier-Ai/README/resolve/main/assets/thumbnail.png
 ---
 
-# Outlier
+<p align="center">
+  <img src="assets/logo.png" alt="Outlier" width="180" />
+</p>
 
-**Patented ternary mixture-of-experts models. Frontier quality on consumer hardware. Apache 2.0.**
+<h1 align="center">Outlier</h1>
 
-Outlier publishes two tracks of models:
+<p align="center"><em>Frontier AI on hardware you own.</em></p>
 
-1. **Consumer tier** — MLX (Apple Silicon) and GGUF (cross-platform) quantized builds ready to run in Ollama, LM Studio, Jan, or the [Outlier desktop app](https://outlier.host).
-2. **Research / server tier** — full-precision ternary-MoE overlays on Qwen 2.5 (V3.2, V3.3) for evaluation and further research.
+<p align="center">
+  <a href="https://huggingface.co/Outlier-Ai/Outlier-Nano-1.7B-MLX-4bit">Nano (1.7B)</a>
+  ·
+  <a href="https://huggingface.co/Outlier-Ai/Outlier-Lite-7B-MLX-4bit">Lite (7B)</a>
+  ·
+  <a href="https://huggingface.co/Outlier-Ai/Outlier-Compact-14B-MLX-4bit">Compact (14B)</a>
+  ·
+  <a href="https://huggingface.co/Outlier-Ai/Outlier-Max-32B-GGUF">Max (32B)</a>
+</p>
 
-## Consumer tier
+---
 
-| Tier | MLX | GGUF | RAM | Target platform |
-|---|---|---|---|---|
-| Nano (1.7B) | [Nano-MLX](https://huggingface.co/Outlier-Ai/Outlier-Nano-1.7B-MLX-4bit) | — | 4 GB+ | Any M-series Mac |
-| **Lite (7B)** | [Lite-MLX](https://huggingface.co/Outlier-Ai/Outlier-Lite-7B-MLX-4bit) | [Lite-GGUF](https://huggingface.co/Outlier-Ai/Outlier-Lite-7B-GGUF) | 8 GB+ | **Default tier** |
-| Compact (14B) | [Compact-MLX](https://huggingface.co/Outlier-Ai/Outlier-Compact-14B-MLX-4bit) | [Compact-GGUF](https://huggingface.co/Outlier-Ai/Outlier-Compact-14B-GGUF) | 16 GB+ | Reasoning, code |
-| Max (32B) | — | [Max-GGUF](https://huggingface.co/Outlier-Ai/Outlier-Max-32B-GGUF) | 32 GB+ | Heaviest workloads |
+## What is Outlier
 
-Fastest path for a Mac user: install the desktop app — https://outlier.host.
+A family of small, fast, local language models designed for the hardware people actually own — Apple Silicon Macs, consumer NVIDIA GPUs, and CPU-only laptops via GGUF.
 
-## Research / server tier (V3.2 family)
+Four tiers from 1.7B to 32B, shipped in three formats:
 
-| Variant | MMLU | Provenance |
+| Format | Use case | Runtime |
 |---|---|---|
-| [Outlier-10B-V3.2](https://huggingface.co/Outlier-Ai/Outlier-10B-V3.2) | ~76% | `[CLAIM]` smoke test |
-| [Outlier-40B-V3.2](https://huggingface.co/Outlier-Ai/Outlier-40B-V3.2) | **77.80%** | `[VERIFIED]` n=14,042 |
-| [Outlier-70B-V3.2](https://huggingface.co/Outlier-Ai/Outlier-70B-V3.2) | **81.49%** | `[VERIFIED]` n=14,042 |
-| [Outlier-150B-V3.2](https://huggingface.co/Outlier-Ai/Outlier-150B-V3.2) | ~83% | `[INCOMPLETE]` provenance |
-
-Per Rule 66: `[VERIFIED]` = complete source file + provenance; `[INCOMPLETE]` = number on disk but fields stripped; `[CLAIM]` = historical smoke-test pending re-run.
-
-## V3.3 preview
-
-- [Outlier-70B-V3.3](https://huggingface.co/Outlier-Ai/Outlier-70B-V3.3) (weights upload pending, Day 18)
-
-## Architecture — plain English
-
-Outlier flagship models use **ternary mixture-of-experts**: weights are {-1, 0, +1} instead of 16-bit floats, so a matrix multiply becomes a stream of additions and subtractions. Combined with **expert paging** (only the experts a given token routes to are resident in RAM; the rest live on disk), Outlier runs a 150B-parameter model on a single Mac Studio at usable speeds.
-
-Three provisional patents filed April 2026 (61 claims total) cover the ternary MoE composition, expert paging, and specialist merging techniques.
-
-## Quick start
+| **MLX 4-bit** | Apple Silicon Macs | Outlier Desktop app, `mlx_lm` |
+| **AutoAWQ 4-bit** | NVIDIA GPUs (datacenter) | vLLM, transformers+awq |
+| **GGUF Q4_K_M / Q5_K_M** | CPU-portable, any OS | llama.cpp, Ollama, LM Studio |
 
-**Mac (fastest):** [outlier.host](https://outlier.host) — download the desktop app.
+## Quickstart — Apple Silicon
 
-**Mac command-line (MLX):**
-```python
+```bash
+pip install mlx-lm
+python -c "
 from mlx_lm import load, generate
-m, t = load("Outlier-Ai/Outlier-Lite-7B-MLX-4bit")
-print(generate(m, t, "What is the capital of France?", max_tokens=200))
+model, tok = load('Outlier-Ai/Outlier-Lite-7B-MLX-4bit')
+print(generate(model, tok, prompt='Hello', max_tokens=100))
+"
 ```
 
-**Any platform (Ollama):**
-```bash
-ollama run hf.co/Outlier-Ai/Outlier-Lite-7B-GGUF:Q4_K_M
-```
+## Quickstart — Ollama
 
-**Server-grade research (Transformers):**
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-m = AutoModelForCausalLM.from_pretrained(
-    "Outlier-Ai/Outlier-40B-V3.2", trust_remote_code=True, torch_dtype="auto"
-)
+```bash
+ollama pull hf.co/Outlier-Ai/Outlier-Lite-7B-GGUF
+ollama run hf.co/Outlier-Ai/Outlier-Lite-7B-GGUF "Hello"
 ```
 
-## Collections
+## Desktop app
 
-- [Outlier Consumer Edition](https://huggingface.co/collections/Outlier-Ai/outlier-consumer-edition-69e2fb4a0df119ea1747275e)
-- [Outlier Server V3.2](https://huggingface.co/collections/Outlier-Ai/outlier-server-v32-69e2fb4b71984614b3c7a4a3)
-- [Outlier Research](https://huggingface.co/collections/Outlier-Ai/outlier-research-69e2fb3a71984614b3c7a279)
+Outlier ships a native Mac desktop app with streaming chat, markdown rendering, session history, model switching, and a full settings panel. See [outlier.host](https://outlier.host) (coming soon) or download from the Releases page.
 
-## Patents
+## Research
 
-Three provisional patents filed April 2026 (61 claims). Non-provisional deadline: April 2027. Patent #3 covers pre-attention expert prediction — empirical scaling to 40B/70B is pending cluster validation.
+Three provisional patents filed April 2026 (61 claims total) on the **Path F+ BTX merge architecture** that produces the tier family from base models. Non-provisional consolidation due April 2027.
 
-## Links
+## License
 
-- Website: https://outlier.host
-- GitHub: https://github.com/Outlier-host/Outlier
-- Desktop app (Mac): https://outlier.host/download (notarization pending)
-- Founders deal: $199 for 2 years — first 100 seats
-
-## Citation
-
-```bibtex
-@misc{outlier2026,
-  author       = {Kerr, Matt},
-  title        = {Outlier: Ternary Mixture-of-Experts for Consumer Hardware},
-  year         = {2026},
-  publisher    = {Hugging Face},
-  howpublished = {\url{https://huggingface.co/Outlier-Ai}}
-}
-```
+Apache 2.0 on all models. Base weights inherit Qwen 2.5's license. Post-training artifacts released under the same terms.
 
-## Roadmap
+## Contact
 
-- **Day 18:** V3.3 weights uploads; MMLU re-verification on 10B-V3.2 and 150B-V3.2; Apple Dev ID enrollment.
-- **Day 20:** Path F+ v3 (code specialist retrain 500 → 2000 steps).
-- **Day 24:** ICML 2026 Workshop submission; public launch.
-- **Later:** True warehouse paging at 150B; shadow-space BTX merge (Claim B); desktop-app notarization.
+[mattkerr09@gmail.com](mailto:mattkerr09@gmail.com) — Matt Kerr, solo founder, Michigan.