Update README.md

README.md

@@ -24,41 +24,22 @@ tags:
   - long-context
 ---
 
-<div align="center">
-
-<a href="https://www.orcarouter.ai">
-  <img src="https://www.orcarouter.ai/orca-logo-classic.png" width="96" alt="OrcaRouter">
-</a>
+<p align="center">
+  <a href="https://www.orcarouter.ai">
+    <img src="https://www.orcarouter.ai/orca-logo-classic.png" width="96" alt="OrcaRouter">
+  </a>
+</p>
 
-# OrcaSAQ2 27B
+<h1 align="center">OrcaSAQ2 27B</h1>
 
-### High-fidelity 3-bit Qwen3.8 for long-horizon agents.
+<p align="center"><strong>High-fidelity 3-bit Qwen3.8 for long-horizon agents.</strong></p>
 
-**54 GB → 12.3 GB · +0.02% PPL · 93.2% Top-1 Agreement · 0.031 KLD · 262K Context**
+<p align="center"><strong>54 GB → 12.3 GB · +0.02% PPL · 93.2% Top-1 Agreement · 0.031 KLD · 262K Context</strong></p>
 
-<p>
-<img src="https://img.shields.io/badge/Checkpoint-12.3_GB-FF6A00?style=for-the-badge">
-<img src="https://img.shields.io/badge/Decoder-3.21_bpw-111111?style=for-the-badge">
-<img src="https://img.shields.io/badge/PPL_Delta-%2B0.02%25-111111?style=for-the-badge">
-<img src="https://img.shields.io/badge/Top--1-93.2%25-111111?style=for-the-badge">
+<p align="center">
+<img src="https://img.shields.io/badge/12.3GB-checkpoint-FF6A00?style=flat-square"><img src="https://img.shields.io/badge/3.21bpw-decoder-111111?style=flat-square"><img src="https://img.shields.io/badge/+0.02%25-PPL-111111?style=flat-square"><img src="https://img.shields.io/badge/93.2%25-Top--1-111111?style=flat-square"><img src="https://img.shields.io/badge/262K-context-111111?style=flat-square">
 </p>
 
-<p>
-<img src="https://img.shields.io/badge/Context-262K-111111?style=flat-square">
-<img src="https://img.shields.io/badge/vLLM-Production_Ready-111111?style=flat-square">
-<img src="https://img.shields.io/badge/MTP-Speculative_Decoding-111111?style=flat-square">
-<img src="https://img.shields.io/badge/Tool_Calling-Yes-111111?style=flat-square">
-<img src="https://img.shields.io/badge/License-Apache_2.0-111111?style=flat-square">
-</p>
-
-**[OrcaRouter](https://www.orcarouter.ai)** ·
-**[Models](https://www.orcarouter.ai/models)** ·
-**[GitHub](https://github.com/Continuum-AI-Corp)** ·
-**[Discord](https://discord.gg/yAh6Tex6kx)** ·
-**[X](https://x.com/OrcaRouter)**
-
-</div>
-
 ---
 
 > ## 27B reasoning. 12.3 GB.
@@ -89,7 +70,7 @@ The resulting checkpoint provides:
 
 ---
 
-# At a glance
+## At a glance
 
 | Metric | BF16 | OrcaSAQ2 |
 |---|---:|---:|
@@ -103,19 +84,13 @@ The resulting checkpoint provides:
 | Mean KLD | — | **0.031** |
 | Context | 262K | **262K** |
 
-<div align="center">
-
-## **4.4× smaller. +0.02% perplexity.**
-
-### The point is not 3-bit.
-
-### **The point is what survives at 3-bit.**
-
-</div>
+<p align="center"><strong>4.4× smaller. +0.02% perplexity.</strong></p>
+<p align="center">The point is not 3-bit.</p>
+<p align="center"><strong>The point is what survives at 3-bit.</strong></p>
 
 ---
 
-# BF16 fidelity
+## BF16 fidelity
 
 All numbers below are measured using **these exact OrcaSAQ2 weights** against the BF16 reference through the same evaluation path.
 
@@ -161,9 +136,9 @@ OrcaSAQ2
 
 ---
 
-# Long-horizon agents
+## Long-horizon agents
 
-## Low-bit fidelity matters more as the horizon grows.
+### Low-bit fidelity matters more as the horizon grows.
 
 Short benchmarks can hide small degradation.
 
@@ -225,7 +200,7 @@ For agent models, both matter.
 
 ---
 
-# Long-horizon performance
+## Long-horizon performance
 
 Long-horizon evaluations should use the same:
 
@@ -242,7 +217,7 @@ Scores produced by materially different agent stacks should not be treated as di
 
 ---
 
-# Architecture
+## Architecture
 
 | | |
 |---|---|
@@ -264,9 +239,9 @@ Scores produced by materially different agent stacks should not be treated as di
 
 ---
 
-# Production serving
+## Production serving
 
-## Up to 90.1 tok/s single-stream on a 16 GB GPU
+### Up to 90.1 tok/s single-stream on a 16 GB GPU
 
 Measured under a **15.7 GiB GPU memory cap**.
 
@@ -302,7 +277,7 @@ For highly batched workloads, benchmark both configurations.
 
 ---
 
-# 27B on a 16 GB GPU
+## 27B on a 16 GB GPU
 
 OrcaSAQ2's checkpoint is **12.3 GB**.
 
@@ -336,7 +311,7 @@ The model architecture supports up to **262K context**.
 
 ---
 
-# Built for agents
+## Built for agents
 
 ### Long-horizon execution
 
@@ -366,9 +341,9 @@ A **12.3 GB** checkpoint designed around practical inference hardware.
 
 ---
 
-# Quickstart
+## Quickstart
 
-## Install
+### Install
 
 ```bash
 pip install -U vllm huggingface_hub
@@ -378,7 +353,7 @@ pip install git+https://github.com/Continuum-AI-Corp/OrcaSAQ2-kernel
 
 ---
 
-## Download
+### Download
 
 ```bash
 hf download orcarouter/OrcaSAQ2-27B \
@@ -387,7 +362,7 @@ hf download orcarouter/OrcaSAQ2-27B \
 
 ---
 
-## Serve with vLLM
+### Serve with vLLM
 
 ```bash
 vllm serve ./OrcaSAQ2-27B \
@@ -401,7 +376,7 @@ vllm serve ./OrcaSAQ2-27B \
 
 ---
 
-# OpenAI-compatible API
+## OpenAI-compatible API
 
 ```python
 from openai import OpenAI
@@ -426,7 +401,7 @@ print(response.choices[0].message.content)
 
 ---
 
-# Recommended sampling
+## Recommended sampling
 
 ```text
 temperature = 1.0
@@ -440,23 +415,19 @@ For agent deployments, benchmark against the actual tool schema, context distrib
 
 ---
 
-# Evaluation philosophy
+## Evaluation philosophy
 
 A low-bit reasoning model should not be judged by checkpoint size alone.
 
 We look at the intersection of:
 
-<div align="center">
-
-### **Footprint × BF16 Fidelity × Capability × Long-Horizon Stability × Serving Performance**
-
-</div>
+<p align="center"><strong>Footprint × BF16 Fidelity × Capability × Long-Horizon Stability × Serving Performance</strong></p>
 
 A useful low-bit model must remain useful after compression.
 
 ---
 
-# Why perplexity alone is not enough
+## Why perplexity alone is not enough
 
 Perplexity is useful and reproducible.
 
@@ -482,7 +453,7 @@ That is why OrcaSAQ2 reports BF16 fidelity metrics alongside downstream and long
 
 ---
 
-# Method
+## Method
 
 OrcaSAQ2 uses a proprietary **sensitivity-aware mixed-precision quantization system** developed by OrcaRouter.
 
@@ -492,7 +463,7 @@ Detailed quantization methodology, calibration strategy, precision allocation an
 
 ---
 
-# Limitations
+## Limitations
 
 - OrcaSAQ2 inherits the capabilities, biases and limitations of Qwen3.8-27B.
 - Quantization is not mathematically lossless.
@@ -506,7 +477,7 @@ Detailed quantization methodology, calibration strategy, precision allocation an
 
 ---
 
-# Open source from OrcaRouter
+## Open source from OrcaRouter
 
 ### [OrcaCode Review](https://github.com/Continuum-AI-Corp/Orca-Code-Review)
 
@@ -520,15 +491,11 @@ Record, replay, fork and debug AI-agent runs.
 
 Self-hosted multi-model AI infrastructure.
 
-<div align="center">
-
-### **Open model. Open harness. Open bill.**
-
-</div>
+<p align="center"><strong>Open model. Open harness. Open bill.</strong></p>
 
 ---
 
-# Citation
+## Citation
 
 ```bibtex
 @misc{qwen38,
@@ -542,7 +509,7 @@ Self-hosted multi-model AI infrastructure.
 
 ---
 
-# License
+## License
 
 **Apache-2.0**
 
@@ -554,16 +521,14 @@ Quantization does not change the underlying license obligations.
 
 ---
 
-<div align="center">
-
-# One Gateway. Every Model.
+<h2 align="center">One Gateway. Every Model.</h2>
 
-**Route Smarter · Ship Safer · Spend Less**
+<p align="center"><strong>Route Smarter · Ship Safer · Spend Less</strong></p>
 
-[Website](https://www.orcarouter.ai) ·
-[Models](https://www.orcarouter.ai/models) ·
-[GitHub](https://github.com/Continuum-AI-Corp) ·
-[Discord](https://discord.gg/yAh6Tex6kx) ·
-[X](https://x.com/OrcaRouter)
-
-</div>
+<p align="center">
+  <a href="https://www.orcarouter.ai">Website</a> ·
+  <a href="https://www.orcarouter.ai/models">Models</a> ·
+  <a href="https://github.com/Continuum-AI-Corp">GitHub</a> ·
+  <a href="https://discord.gg/yAh6Tex6kx">Discord</a> ·
+  <a href="https://x.com/OrcaRouter">X</a>
+</p>