feat: add cluster collection and quality tiers to featured models

[ashokgelal]

Problem

Cluster Advisor in Msty Nexus (cloudstack-llc/msty-nexus#583) ranks models for a Mac cluster. It reads the featured list's cluster collection and a quality tier per model. Without them, it falls back to the featured MLX models, which are mostly small. The result is weak picks on big clusters.

Solution

• New cluster collection with 10 MLX models.
• New optional tier on every model: frontier, excellent, strong, or good.
• "coding" in tags on 15 models that are good at code.
• schema_version stays 1. Older Nexus builds ignore the new fields.
• Existing rows are unchanged except for the new tier and tags.

Cluster picks

Model	Tier	MLX repository	Download
Muse Spark Open 816B	frontier	mlx-community/Muse-Spark-Open-816B-A42B-4bit	532.8 GB
GLM-5.3	frontier	mlx-community/GLM-5.3-4bit	418.3 GB
DeepSeek V3.2	excellent	mlx-community/DeepSeek-V3.2-4bit	378.1 GB
Qwen3 Coder 480B	excellent	mlx-community/Qwen3-Coder-480B-A35B-Instruct-4bit	270.1 GB
Qwen3.5 397B	excellent	mlx-community/Qwen3.5-397B-A17B-4bit	223.9 GB
GLM-4.7	excellent	mlx-community/GLM-4.7-4bit	198.6 GB
MiniMax M2.7	excellent	mlx-community/MiniMax-M2.7-4bit	128.7 GB
Qwen3.5 122B (existing)	strong	mlx-community/Qwen3.5-122B-A10B-4bit	69.6 GB
gpt-oss 120B (existing)	strong	mlx-community/gpt-oss-120b-MXFP4-Q8	63.4 GB
GLM-4.5 Air	strong	mlx-community/GLM-4.5-Air-4bit	60.2 GB

The builder read every repository and size from the Hugging Face API (the sum of the repository's files). It also checked that each model_type is a family the cluster worker can split (internal/cluster/model_capabilities.go), and that no pick needs custom Python code to load.

I used the newest version of each family: GLM-4.7 instead of 4.6, MiniMax M2.7 instead of M2, Qwen3.5 397B instead of Qwen3 235B. GLM-4.5 Air and Qwen3 Coder 480B are older, but they fill a size or a coding gap.

Left out

• Kimi K2.5 / K2.6 / K2.7 Code: the cluster can split them (kimi_k25), but their tokenizer needs trust_remote_code (tokenization_kimi.py). Nexus never passes that flag, so I expect the load to fail. They can be added after we decide on that flag.
• Newer families the cluster cannot split yet: DeepSeek V4 (deepseek_v4), Kimi K3 (kimi_k3), MiniMax M3 (minimax_m3_vl), GLM-5.3 Flash (glm5_next), Qwen3.8 Flash Next (qwen4_exp), Mistral Large 3 (mistral_large3), Nemotron 3 Ultra (nemotron_h).

How tiers were chosen

A tier is a curated rating of general chat and reasoning quality, compared with the open models available now:

• frontier: the best open models, close to hosted frontier models.
• excellent: one step below: the previous generation's flagships and the top coding models.
• strong: capable mid-size models, roughly 20B to 130B.
• good: small models, and models with one narrow job (OCR, embeddings). These must never rank above a chat model.

Tiers are relative, so we must review them when new models are released. The README now documents tier, "coding", and the cluster collection.

Test

• PR #583's Go parser (featuredmodels.parse) reads the file. It returns 9 collections, and all 10 cluster picks have their tier, tags, and an MLX install.
• node --test scripts/*.test.mjs and node scripts/verify.mjs pass.
• A script confirmed that the existing rows and collections are byte-identical apart from tier and tags.

Notes

• Older Nexus builds show "For a Mac cluster" as a new row on the featured models page.
• Do not merge before the maintainer approves. A merge to main publishes the file.
• Builder change: cloudstack-llc/msty-nexus PR (linked below). I did not publish a full rebuild. Ollama now lists some sizes as ranges ("6.6GB - 7.6GB"), and the builder cannot read them, so 8 existing Ollama installs would lose size_bytes.