huggingface-scout

name	huggingface-scout
description	Search HuggingFace Hub for models, datasets, and training resources. Use for model discovery, dataset search, release monitoring, and candidate evaluation.
metadata	{"author":"Fae","version":"1.0.0"}
tags	["huggingface","models","datasets","training","self-improvement"]

HuggingFace Scout

Search HuggingFace Hub for models, datasets, and training resources to improve Fae.

Commands

search_models — Find models

{"query": "Qwen3.5 MLX", "limit": 5}
{"query": "MLX 4bit", "filter_tags": ["text-generation"], "sort": "downloads"}

search_datasets — Find datasets

{"query": "instruction tuning", "limit": 5}
{"query": "preference DPO", "filter_tags": ["en"]}

check_new_releases — Monitor new models

{"since": "2026-03-01"}
{"since": "2026-03-01", "authors": ["mlx-community", "Qwen"]}

evaluate_candidate — Assess compatibility

{"model_id": "Brooooooklyn/Qwen3.5-9B-unsloth-mlx"}

Checks: architecture, context length, MLX support, license, downloads, parameter count.

Fae's Current Stack

Role	Model	Params
Operator	Qwen3.5-2B	2B 4-bit MLX (OptiQ mixed-precision)
Operator	Qwen3.5-4B	4B 4-bit MLX (OptiQ mixed-precision)
Operator	Qwen3.5-9B	9B 4-bit MLX
Operator	Qwen3.5-35B-A3B	35B total / 3B active per token MoE
STT	Qwen3-ASR-1.7B	1.7B 4-bit
TTS	Kokoro-82M	82M float32
VLM	Qwen3-VL-4B/8B	4-8B
Speaker	ECAPA-TDNN	~6M CoreML

Fae uses a single local model at a time, selected by RAM tier:

• Qwen3.5-2B: Any Mac with 8+ GB RAM
• Qwen3.5-4B: Macs with 16+ GB RAM
• Qwen3.5-9B: Macs with 24+ GB RAM
• Qwen3.5-35B-A3B: Macs with 32+ GB RAM

When evaluating candidate models, prioritise MLX-native quantisations (4-bit, 8-bit) with context windows of 32K+ tokens.