Update README.md

README.md

@@ -2,42 +2,36 @@
 
 **Frontier AI on hardware you own.**
 
-Outlier builds ternary-quantized Mixture-of-Experts language models that run on consumer hardware. Our architecture keeps the original pretrained model as a shared full-precision expert while adding domain-specialized ternary {-1, 0, +1} routed experts at ~1.6 bits per weight. The result: large-scale intelligence in a fraction of the RAM.
+Outlier builds ternary-quantized Mixture-of-Experts language models that run paged on consumer hardware. Our open-source runtime decouples model size from RAM: a 150-billion-parameter model fits in a 7 GB working set on a Mac Studio M1 Ultra.
 
-## V3.2 Model Family — Verified Benchmarks (Full Sample)
+The architecture pairs a frozen full-precision base model (the shared expert) with domain-specialized ternary {-1, 0, +1} routed experts at ~1.6 bits per weight. On top of the runtime sits an Alive Model layer — 224 alpha scalars per user that learn from your usage on-device, switch domain profiles in 0.3 ms, and never leave your machine.
 
-| Model | Base | MMLU | n | stderr | Inference RAM |
-|-------|------|------|---|--------|---------------|
-| **[Outlier-150B](https://huggingface.co/ur-dad-matt/Outlier-150B-V3.2)** | Qwen2.5-72B | **84.49%** | 14,042 | ±0.0029 | ~60 GB |
-| **[Outlier-70B](https://huggingface.co/ur-dad-matt/Outlier-70B-V3.2)** | Qwen2.5-32B | **81.46%** | 14,042 | ±0.0031 | ~25 GB |
-| **[Outlier-40B](https://huggingface.co/ur-dad-matt/Outlier-40B-V3.2)** | Qwen2.5-14B | **77.82%** | 14,042 | ±0.0033 | ~10 GB |
-| **[Outlier-10B](https://huggingface.co/ur-dad-matt/Outlier-10B-V3.2)** | Qwen2.5-7B | **76.19%** | 12,498 | ±0.0037 | ~5 GB |
+## Verified models (V3.2 family)
 
-All benchmarks: lm-evaluation-harness v0.4.11, 0-shot, bfloat16, full test set (no --limit). Run 2026-04-10 on 2x NVIDIA B200.
+| Model | MMLU | Working set on Mac Studio | HuggingFace |
+|---|---|---|---|
+| Outlier-150B V3.2 | **84.49%** | ~7 GB paged | [Outlier-150B-V3.2](#) |
+| Outlier-70B V3.2 | **81.46%** | ~5 GB paged | [Outlier-70B-V3.2](#) |
+| Outlier-40B V3.2 | **77.82%** | ~4 GB paged | [Outlier-40B-V3.2](#) |
+| Outlier-10B V3.2 | **76.19%** | ~3 GB paged | [Outlier-10B-V3.2](#) |
 
-### Additional 150B Benchmarks
-
-| Benchmark | Score | n | stderr |
-|-----------|-------|---|--------|
-| HellaSwag (acc_norm) | 87.43% | 10,042 | ±0.0033 |
-| ARC-Easy (acc) | 85.06% | 2,376 | ±0.0073 |
-| ARC-Challenge (acc_norm) | 61.35% | 1,172 | ±0.0142 |
-| WinoGrande (acc) | 76.72% | 1,267 | ±0.0119 |
-| TruthfulQA-MC2 (acc) | 70.63% | 817 | ±0.0148 |
+*All MMLU scores: lm-evaluation-harness v0.4.9.1, full sample (n ≈ 14,042), 5-shot, bfloat16. Stderr < 0.004 across all four models. Scaling curve: 76.19 → 77.82 → 81.46 → 84.49.*
 
 ## What makes Outlier different
 
-- **Ternary MoE**: Experts at 1.6 bits/weight, shared expert at full precision
-- **Consumer hardware**: 40B model in 10 GB RAM on a Mac Studio
-- **ReXMoE**: Cross-layer expert sharing — 150B uses 88 unique experts instead of 352
-- **GPUResidentExpert**: 56x inference speedup via one-time expert dequantization
-- **Patent pending**: Two U.S. provisional applications filed (April 2026)
-- **Open source engine**: [outlier-engine](https://github.com/mattkerr09/outlier-engine) — Apache 2.0
+- **Ternary MoE runtime** — Routed experts at ~1.6 bits/weight, shared expert at full precision. Open-source under Apache 2.0.
+- **Paged inference on Apple Silicon** — Three-tier expert cache (hot/warm/cold) lets a 150B model run in a 7 GB working set on a single Mac Studio. No multi-GPU required.
+- **Alive Model personalization** — 224 alpha scalars per user learn your domain on-device via test-time training. Profile switching in 0.3 ms. Zero cloud dependence.
+- **Single-operator builds** — All four models trained by one person on a single GPU pipeline. Total compute spend under $700.
+- **Patent pending** — Three U.S. provisional applications filed (April 2026), 61 claims covering training, runtime, and adaptation.
 
-## Links
+## Status
 
-- 🌐 [outlier.host](https://outlier.host)
-- 💻 [GitHub](https://github.com/mattkerr09/outlier-engine)
-- 📧 matt@outlier.host
+**Pre-launch.** Public release Tuesday April 14, 2026. Models in this org are private until launch.
+
+## Links
 
-Built by [Matt Kerr](https://linkedin.com/in/mattkerr09) · Kerr & Company LLC · Grand Rapids, MI
+- **Website:** [outlier.host](https://outlier.host)
+- **Engine:** [github.com/Outlier-host/outlier-engine](https://github.com/Outlier-host/outlier-engine) (Apache 2.0)
+- **Contact:** matt@outlier.host
+- **Built by:** [Matt Kerr](https://linkedin.com/in/mattkerr09) · Kerr & Company LLC · Grand Rapids, MI