Add Primitive organization card

Brand banner, positioning, public Qwen3.8-27B metrics, models index. Contains no data from private repos.

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/org-banner.png filter=lfs diff=lfs merge=lfs -text
+assets/qwen-metrics.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,10 +1,116 @@
 ---
 title: README
-emoji: 🦀
-colorFrom: green
-colorTo: blue
+emoji: 🟧
+colorFrom: yellow
+colorTo: red
 sdk: static
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+<p align="center">
+  <img src="assets/org-banner.png" alt="Primitive — inference economics for production LLM systems" width="100%">
+</p>
+
+<p align="center">
+  <img src="https://img.shields.io/badge/focus-inference%20economics-F98424?style=flat-square" alt="Inference economics">
+  <img src="https://img.shields.io/badge/runtimes-vLLM%20·%20ONNX-F98424?style=flat-square" alt="vLLM and ONNX">
+  <img src="https://img.shields.io/badge/hardware-Blackwell%20·%20A100%20·%20CPU-F98424?style=flat-square" alt="Blackwell, A100, CPU">
+  <img src="https://img.shields.io/badge/weights-Apache--2.0-lightgrey?style=flat-square" alt="Apache 2.0">
+</p>
+
+---
+
+## Serving LLMs well is two problems
+
+**Which model should answer this request?** and **how cheaply can that model
+run?** Most teams solve neither — they pin one frontier model to every route and
+pay frontier prices on the easy majority of their traffic.
+
+We build the two pieces that fix that, and we publish the measurements.
+
+<table>
+<tr>
+<td width="50%" valign="top">
+
+### 🔀 Route
+
+A small, fast model that reads the prompt **before** you spend anything on it,
+then ranks which LLM is actually worth calling. Runs on CPU next to your API
+server — no GPU, no PyTorch, no extra service to operate.
+
+*`llm-router-lite-80m-onnx` — private beta.*
+
+</td>
+<td width="50%" valign="top">
+
+### 🗜️ Compress
+
+Mixed-precision quantization that shrinks a checkpoint and speeds up serving
+**without quietly trading away accuracy** — and that still loads on the GPUs
+you already have, not just the newest ones.
+
+*`Qwen3.8-27B-mixed-NVFP4-FP8` — public, open weights.*
+
+</td>
+</tr>
+</table>
+
+---
+
+## Latest release: Qwen3.8-27B, 2.5× smaller, faster, accuracy intact
+
+<p align="center">
+  <img src="assets/qwen-metrics.png" alt="Qwen3.8-27B mixed NVFP4/FP8: 21 GiB vs 51.8 GiB BF16, accuracy 88.5 vs 88.6 BF16, 1134 tok/s vs 560 tok/s at concurrency 32" width="100%">
+</p>
+
+| config | size | accuracy | tok/s @ conc 32 | tok/s @ conc 1 |
+|---|---|---|---|---|
+| BF16 | 51.8 G | 88.6 | 559.6 | 26.3 |
+| Qwen FP8 | 28.8 G | 87.6 | 884.0 | 45.6 |
+| **primitive mixed NVFP4/FP8** | **~21 G** | **88.5** | **1134.5 (2.0×)** | **66.4 (2.5×)** |
+
+Measured on a 1,170-sample mixed suite. Accuracy is statistically tied with BF16
+(run-to-run spread ~±0.8). Serves on stock vLLM on **Blackwell** via native
+FP4/FP8 kernels and on **Ampere/A100** via the Marlin W·A16 fallback — 88.1
+accuracy, 790 tok/s @ concurrency 32 on an A100 40 GB. Speculative decoding
+works on the quant: up to **+101%** with a DSpark drafter.
+
+```bash
+vllm serve primitive-ai/Qwen3.8-27B-mixed-NVFP4-FP8 \
+  --reasoning-parser qwen3 \
+  --speculative-config '{"method":"mtp","num_speculative_tokens":2}'
+```
+
+---
+
+## How we publish
+
+* **Numbers or nothing.** Every figure on our model cards comes from a held-out
+  suite and is reproducible from files in the repo. Where a metric is relative,
+  we say what it's relative to.
+* **We document what breaks.** Our cards carry the failure modes, not just the
+  wins — which head dtype makes an A100 load succeed, which upstream config
+  field crashes vLLM, which optimisation we *rejected* and why.
+* **No silent quality loss.** A compression or routing win that changes answers
+  isn't a win. When a cheaper option moved outputs too much, we don't ship it —
+  and we say so on the card.
+* **Runs on hardware you have.** Ampere fallbacks for the quants, CPU-only for
+  the router. Nothing here assumes you just bought a Blackwell rack.
+
+---
+
+## Models
+
+| repo | what it does | status |
+|---|---|---|
+| [`Qwen3.8-27B-mixed-NVFP4-FP8`](https://huggingface.co/primitive-ai/Qwen3.8-27B-mixed-NVFP4-FP8) | Mixed NVFP4/FP8 weights-only quant of Qwen3.8-27B. vLLM-native, Blackwell + A100, MTP head preserved for speculative decoding. | Public |
+| `llm-router-lite-80m-onnx` | CPU-only router: reads a chat `messages` array, ranks which LLM to call. ONNX, no torch or transformers. | Private beta |
+
+All published weights are Apache-2.0. Derivative models carry upstream
+attribution and a `NOTICE` file.
+
+<p align="center">
+  <br>
+  <img src="assets/primitive-logo.png" alt="Primitive" width="34"><br>
+  <sub><b>primitive</b> · inference economics for production LLM systems</sub>
+</p>