Model card: conversion provenance, architecture, MTP head pointer, challenge role

README.md

@@ -2,6 +2,46 @@
 language: en
 library_name: mlx
 pipeline_tag: text-generation
+license: apache-2.0
+base_model: Qwen/Qwen3.8-27B
 tags:
 - mlx
+- 4-bit
+- qwen3_5
 ---
+
+# Qwen3.8-27B-4bit
+
+A 4-bit MLX conversion of [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B)
+(revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`), quantized affine 4-bit with
+group size 64 under a pinned **mlx 0.32.0** toolchain via `mlx_lm.convert`.
+
+## Architecture
+
+- 64 layers on a 4-layer hybrid repeat: three linear-attention layers, then one
+  full-attention layer (`full_attention_interval: 4`)
+- hidden size 5120, 24 attention heads / 4 KV heads, head dim 256
+- vocabulary 248,320
+- `model_type: qwen3_5`
+
+## Multi-token-prediction head
+
+The upstream checkpoint ships a native MTP (multi-token-prediction) head as 15
+`mtp.*` tensors. Those tensors are **not** in this repository — they are published
+separately, in their released bf16 dtype, as
+[EigenLabs/Qwen3.8-27B-MTP-bf16](https://huggingface.co/EigenLabs/Qwen3.8-27B-MTP-bf16).
+
+## Role
+
+This is the pinned reference checkpoint for the Qwen 3.8 native-MTP
+speculative-decode challenge track (`qwen3.8-27b-mtp-v1`) at
+[Layr-Labs/qwen-3.8-mtp-challenge](https://github.com/Layr-Labs/qwen-3.8-mtp-challenge).
+The track pins this repository by revision; scoring artifacts are byte-verified
+against it.
+
+## Usage
+
+```bash
+pip install mlx-lm
+mlx_lm.generate --model EigenLabs/Qwen3.8-27B-4bit --prompt "Hello"
+```