Model Mania

Local-first open-weight model catalog for RTX 5070 Ti 16GB + 31.4 GB RAM

Purpose

Single source of truth for every free/open-weight model that can run locally on a standalone 5070 Ti. Organized by execution branch for runtime routing in AI Tracker / Hermes workflows.

Hardware Baseline

• GPU: RTX 5070 Ti 16GB VRAM
• System RAM: 31.4 GB
• Ollama runtime: localhost:11434
• Verified env: OLLAMA_NUM_GPU=999, CUDA_VISIBLE_DEVICES=0

Branches

• gpu-only/ — weights fit fully in VRAM
• hybrid/ — partial GPU offload + CPU RAM layers
• cpu-fallback/ — no GPU offload
• embeddings/ — embedding/retrieval
• vision/ — multimodal image-capable
• audio/ — speech/audio

Verification Rule

No entry is accepted without one of:

• local Ollama manifest size + runtime evidence
• official model card size
• documented quant size from a verifiable source

Sizing provenance tags

• local_verified: true — confirmed from local Ollama manifest + runtime
• local_verified: false + verified_date: null — size from Ollama library page or public estimate only; recheck before first pull

Workflow

1. Identify candidate
2. Record exact size from manifest/model card
3. Test on host
4. Write catalog entry
5. Commit with evidence path