Picking a model

What the Plan tab does

The Plan tab is a catalog browser over the open-model ecosystem (Llama, Qwen, Mistral, Phi, DeepSeek, Gemma, and friends). For each model it knows:

• Architecture fields (params, layers, KV heads, hidden size) — used by the Hardware tab to compute exact VRAM.
• License (MIT, Apache, Llama community, etc.) — relevant for commercial use.
• Quant variants published by the community on HuggingFace — determines what files Blueprint can actually pull onto your disk.
• Popularity rank + capability tags (structured output, multilingual, long-context-proven) — surface filters in the side panel.

You filter, browse, click into a model, then continue to Hardware to size it.

What to look for

In rough priority order:

1. Params vs your VRAM budget. A 7B model in Q4 fits in 6-8 GB. A 70B model in Q4 needs ~40-50 GB. The Hardware tab does the exact math; rough rule for Q4: params * 0.7 GB.
2. Quant options published. Some models only have Q4 and Q5 GGUFs on HF; if you want Q8 for accuracy, check the model's quantOptions field in the detail pane before committing.
3. License. "Open weights" isn't the same as "open license." Llama 3 has a community license with use restrictions. Qwen, Mistral, Phi are mostly Apache. Read the license if you're shipping commercial.
4. Capability tags. If you need JSON output, filter to structuredOutput: true — most small models will hallucinate JSON keys without a trained-for-structure backbone.
5. Recency. The catalog ships with a popularityRank field but skews to launch-week traffic. Cross-reference with HF trending or the LMArena leaderboard for current quality signals.

Help-me-choose

The "Help me choose" button surfaces a guided picker: hardware budget, workload type, latency tolerance. It narrows the catalog to a shortlist and explains why each candidate ranks where it does. Good first stop if you don't already have a model in mind.

Common pitfalls

• Picking by params alone. A 70B Llama-3 in Q4 doesn't beat a well-tuned 32B Qwen-2.5 in Q4 on most benchmarks. Architecture and training data matter more than raw parameter count.
• Ignoring the quant the model was actually trained against. Some models (Phi-4 in particular) are quant-sensitive — Q3 results are noticeably worse than Q4. Default to Q4 unless you've validated.
• Forgetting context length. The detail pane shows maxContext, but VRAM scales with the context you actually allocate. A 32k-context Q4 7B uses way more VRAM than the same model at 4k.