kaun banega tokenpati, screen wireframe. Build to this, then argue with it.
Terminal width assumed 100. Everything must degrade to 80.

 ██╗  ██╗ █████╗ ██╗   ██╗███╗   ██╗    ██████╗  █████╗ ███╗   ██╗███████╗ ██████╗  █████╗
 (pyfiglet, gradient left to right, accent -> green)
                         k a u n   b a n e g a   T O K E N P A T I
                               "kitna deti hai?"  (dim, italic)

┌─ HOT SEAT ──────────────────────────────────────────────────────────────────────────────┐
│  chip       Apple M3 Max            backend     mlx  (llama.cpp-metal fallback)         │
│  memory     36 GB unified           bandwidth   300 GB/s                                │
│  gpu cores  30                      os reserve  4 GB                                    │
└─────────────────────────────────────────────────────────────────────────────────────────┘

 budgeting for 8,192 context  (--context to change)

 #   contestant          quant    memory  [weights|kv|free]                    tok/s   ctx max   verdict
 1   Qwen2.5 7B          Q8_0     ████████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░    34      32k       daudega
 2   Llama 3.1 8B        Q8_0     ████████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░    31      64k       daudega
 3   Gemma 3 12B         Q6_K     ████████████████░░░░░░░░░░░░░░░░░░░░░░░░    22      32k       chalega
 4   Qwen3 32B           Q4_K_M   ████████████████████████████░░░░░░░░░░░░     9      16k       chalega
 5   Llama 3.3 70B       Q4_K_M   ██████████████████████████████████████████    -       -        ghare jake sutti babu

 bar colours: weights green, kv amber, free dim. sutti rows: whole bar red.
 verdict colours: daudega green bold, chalega amber, sutti red.

┌─ FINAL ANSWER ──────────────────────────────────────────────────────────────────────────┐
│                                                                                         │
│   Qwen2.5 7B  ·  Q8_0  ·  mlx  ·  ~34 tok/s  ·  up to 32k context                       │
│                                                                                         │
│   pip install mlx-lm                                                                    │
│   mlx_lm.generate --model mlx-community/Qwen2.5-7B-Instruct-8bit --prompt "hi"          │
│                                                                                         │
└─────────────────────────────────────────────────────────────────────────────────────────┘
   (accent border, brightest thing on screen)

Later, maybe:
  - speedometer arc next to the final answer
  - a "phone a friend" flag that explains why a model got its verdict
  - --serve, phase 2
