chore(cards): optimize tags + README for discoverability (2026-05-03)

README.md

@@ -1,72 +1,180 @@
 ---
+language:
+  - en
+  - zh
+  - fr
+  - es
+  - pt
+  - de
+  - it
+  - ru
+  - ja
+  - ko
+  - ar
+  - vi
+  - th
+  - nl
+  - pl
 license: apache-2.0
-base_model: mlx-community/gemma-4-26b-a4b-it-4bit
+library_name: mlx
+base_model: google/gemma-4-26b-a4b-it
 tags:
-- mlx
-- gemma
-- moe
-- 4bit
-language:
-- en
-
+  - mlx
+  - safetensors
+  - gemma4
+  - gemma
+  - gemma-4
+  - moe
+  - mixture-of-experts
+  - 4bit
+  - 4-bit
+  - quantized
+  - apple-silicon
+  - mac
+  - macos
+  - metal
+  - m1
+  - m2
+  - m3
+  - m4
+  - mlx-lm
+  - local-llm
+  - on-device
+  - edge-ai
+  - offline
+  - outlier
+  - outlier-app
+  - thinking
+  - reasoning
+  - function-calling
+  - tool-use
+  - text-generation
+  - conversational
+  - instruct
+  - chat
+  - multilingual
+  - mmlu-verified
 pipeline_tag: text-generation
+
+model-index:
+- name: Outlier-Ai/Outlier-Quick-26B-MLX-4bit
+  results:
+  - task:
+      type: text-generation
+      name: Text Generation
+    dataset:
+      name: MMLU (stratified n=300)
+      type: cais/mmlu
+      config: all
+      split: test
+    metrics:
+    - type: acc
+      name: accuracy
+      value: 0.7933
+      verified: false
+  - task:
+      type: text-generation
+      name: Text Generation
+    dataset:
+      name: HumanEval
+      type: openai_humaneval
+      split: test
+    metrics:
+    - type: pass@1
+      name: pass@1
+      value: 0.1280
+      verified: false
 ---
 
-# Outlier Quick (26B MoE / 4B active)
+# Outlier Quick — 26B MoE / 4B active (MLX-4bit)
 
-The Outlier **Quick** tier ships in the Outlier desktop app v1.7+. This
-repo is a re-publish of [`mlx-community/gemma-4-26b-a4b-it-4bit`](https://huggingface.co/mlx-community/gemma-4-26b-a4b-it-4bit)
-under the Outlier-Ai org, with σ-qualified bench provenance attached.
+> **Thinking-mode reasoning on any 16 GB Mac. Fast/light MoE tier in the [Outlier](https://outlier.host) lineup.**
 
-Quick sits between Lite (9B dense) and Core (27B dense) in the Outlier
-lineup — a fast/light tier with knowledge density of a small dense model
-plus thinking-mode reasoning passes. It is **not** a code-tier substitute
-(see HumanEval below); for coding workflows prefer Core or Code.
+**Outlier Quick** is a 4-bit MLX conversion of [Google Gemma 4 26B-a4b-it](https://huggingface.co/google/gemma-4-26b-a4b-it) — a Mixture-of-Experts model with 26B total parameters but only **4B active per token**. It runs at ~14.6 tok/s on M1 Ultra and fits in 16 GB unified RAM, making it one of the fastest thinking-mode models available on Apple Silicon.
 
-- **Base model:** [`google/gemma-4-26b-a4b-it`](https://huggingface.co/google/gemma-4-26b-a4b-it)
-- **Quantization:** mlx-community 4-bit MLX (15.6 GB on disk, 3 safetensors)
-- **Architecture:** Gemma 4 Mixture-of-Experts, 26B total / 4B active per token
-- **Active RAM at inference:** ~15.4 GB (smoke, M1 Ultra 64 GB UMA)
-- **First-token latency (cold load):** 1.22 s (smoke, n=1)
-- **License:** Apache 2.0 (inherited from base)
+Ships in the **[Outlier desktop app](https://outlier.host)** (v1.7+) as the "Quick" tier — positioned between Lite (9B dense) and Core (27B dense).
 
-## σ-qualified evals (Rule 66 5-field provenance)
+---
 
-| Task | Metric | n | 95% CI | Timestamp (UTC) | Notes |
-|---|---|---|---|---|---|
-| MMLU | 0.7933 | 300 | [0.7439, 0.8353] | 2026-04-30T17:50:59.463658+00:00 | Wilson 95% CI |
-| HumanEval pass@1 | 0.1280 | 164 | [0.0853, 0.1878] | 2026-04-30T19:19:29.402729+00:00 | Wilson 95% CI |
+## At a glance
 
-**Sampling note:** MMLU was run on a stratified-by-subject n=300 sample
-because Gemma 4 26B IT is a thinking-mode model (~16-30 s/q wall). Full
-n=14042 would require 8-25 hours on M1 Ultra. The n=300 stratified
-sample covers all 57 MMLU subjects with ±5.7 pp Wilson 95% CI at p=0.5
-(±5.0 pp at p=0.7), which is sufficient resolution for tier-decision
-purposes and honest under Rule 78 (smoke ≠ perf claim — n is reported).
+| Property | Value |
+|---|---|
+| Architecture | Gemma 4 MoE, hybrid sliding-window attention |
+| Total params | 26B (4B active per token) |
+| Quantization | MLX 4-bit (15.6 GB on disk) |
+| Min RAM | **16 GB** unified memory |
+| Speed | **14.6 tok/s** (M1 Ultra, MLX-4bit) |
+| Context (default) | 32K (native: 256K) |
+| Thinking mode | ✅ |
+| License | Apache 2.0 |
 
-HumanEval pass@1 was run over the full n=164 problems.
+---
 
-## Verifying
+## Verified benchmarks (σ-qualified)
 
-```python
-from huggingface_hub import snapshot_download
-local = snapshot_download("Outlier-Ai/Outlier-Quick-26B-MLX-4bit")
+| Task | Metric | n | 95% CI | Date |
+|---|---|---|---|---|
+| MMLU | **0.7933** | 300 (stratified) | [0.744, 0.835] | 2026-04-30 |
+| HumanEval pass@1 | **0.128** | 164 | [0.085, 0.188] | 2026-04-30 |
+
+> ⚠️ **Coding note:** HumanEval pass@1 = 0.128. Quick is **not** a code-generation substitute. For coding use [Outlier Core](https://huggingface.co/Outlier-Ai/Outlier-Core-27B-MLX-4bit) (HumanEval 0.866) or [Outlier Code](https://huggingface.co/Outlier-Ai/Outlier-Code-27B-MLX-4bit). Quick's strength is knowledge-density + reasoning on hard non-code queries.
+
+MMLU stratified sample covers all 57 subjects (n=300 due to Gemma 4's thinking-mode latency ~16–30 s/q on M1 Ultra). Wilson 95% CI ±5.7 pp at p=0.5. HumanEval run over full n=164 problems.
+
+---
+
+## Quick start
 
-# In Python via mlx_lm:
-from mlx_lm import load, generate
-model, tokenizer = load(local)
-print(generate(model, tokenizer, "The capital of France is", max_tokens=32))
+```bash
+pip install mlx-lm
+mlx_lm.generate --model Outlier-Ai/Outlier-Quick-26B-MLX-4bit \
+  --prompt "Explain the difference between MoE and dense transformers." \
+  --max-tokens 512
 ```
 
+```python
+from mlx_lm import load, stream_generate
+
+model, tokenizer = load("Outlier-Ai/Outlier-Quick-26B-MLX-4bit")
+for chunk in stream_generate(
+    model, tokenizer,
+    "What's the most efficient sorting algorithm for nearly-sorted arrays?",
+    max_tokens=256
+):
+    print(chunk.text, end="", flush=True)
+```
+
+---
+
+## Where it fits in the Outlier lineup
+
+| Tier | Params | Speed | Min RAM | Best for |
+|---|---|---|---|---|
+| [Nano](https://hf.co/Outlier-Ai/Outlier-Nano-4B-MLX-4bit) | 4B dense | 71.7 tok/s | 6 GB | Fastest iteration |
+| [Lite](https://hf.co/Outlier-Ai/Outlier-Lite-9B-MLX-4bit) | 9B dense | 53.4 tok/s | 12 GB | Everyday chat |
+| **Quick** ← you are here | 26B MoE / 4B active | **14.6 tok/s** | **16 GB** | Reasoning on 16 GB Mac |
+| [Core](https://hf.co/Outlier-Ai/Outlier-Core-27B-MLX-4bit) | 27B dense | 20.7 tok/s | 24 GB | Best quality/speed |
+| [Code](https://hf.co/Outlier-Ai/Outlier-Code-27B-MLX-4bit) | 27B dense | 20.7 tok/s | 24 GB | HumanEval 0.866 |
+| [Vision](https://hf.co/Outlier-Ai/Outlier-Vision-35B-A3B-MLX-4bit) | 35B MoE / 3.6B active | ~61 tok/s | 24 GB | Image + text |
+
+---
+
+## Technical details
+
+- **Architecture:** `Gemma4ForConditionalGeneration`, 30 layers, hybrid sliding-window attention
+- **Routing:** top-k MoE (4B active from 26B total)
+- **Quantization:** mlx-community 4-bit, 15.61 GB on disk (3 safetensors)
+- **Native context:** 256K via hybrid sliding-window attention
+- **Vision tower:** present in source but not used in text-only inference path
+
 ## Provenance
 
-- Source repo: `mlx-community/gemma-4-26b-a4b-it-4bit` (mlx-community)
-- Re-published by: Outlier-Ai
-- Sprint: `OUTLIER-V17-MAC-ONLY-DAY32-001`
-- Commit: see [`sprints/v17_mac_only/PROVENANCE.md`](https://github.com/Outlier-Ai/outlier-tauri/blob/main/sprints/v17_mac_only/PROVENANCE.md)
+- **Source:** [`mlx-community/gemma-4-26b-a4b-it-4bit`](https://huggingface.co/mlx-community/gemma-4-26b-a4b-it-4bit)
+- **Base model:** [`google/gemma-4-26b-a4b-it`](https://huggingface.co/google/gemma-4-26b-a4b-it)
+- **Published by:** [Outlier-Ai](https://huggingface.co/Outlier-Ai)
+- **Benchmark sprint:** `OUTLIER-V17-MAC-ONLY-DAY32-001` (2026-04-30)
 
 ## License
 
-Apache 2.0 — see [LICENSE](https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md)
-inherited from `mlx-community/gemma-4-26b-a4b-it-4bit` and ultimately from
-`google/gemma-4-26b-a4b-it`.
+Apache 2.0 — inherited from `google/gemma-4-26b-a4b-it`.