Federico

Lesto, a new MLX-based inference engine for @googlegemma 4.

My personal local AI experiment.

The fastest of the 6 MLX engines I tested on a Mac.

Based on a stripped down MLX-Serve by @ddalcu + custom optimizations and optionally with its own optimized checkpoint.

Reminder for myself: local inference stack for DeepSeek V4.1 Flash 1017 (an Opus 5ish OS model)
Approx. total: ~$4,600–$5,200

GPU: ASRock Arc Pro B70 32GB - $1,300 (Newegg)
RAM: NEMIX 256GB (8x32GB) DDR4-3200 ECC - ~$2,370 (Newegg)
SSD (Engram/model): WD Blue SN5000 4TB NVMe - ~$480 (Amazon)
CPU: AMD Ryzen 9 9900X / Intel Core Ultra 7 equivalent - ~$450–550 (Newegg)
Motherboard: X870E / Z890 board (256GB support + PCIe 5.0) - ~$250–350 (Newegg)
PSU: 850W 80+ Gold - ~$120 (Newegg)
Case: Mid-tower - ~$90 (Newegg)

The game of life.

Realize and accept that:
- life is a game
- you have no other option than to play it
- there's no official rules manual (sorry Jordan Peterson)

Choose your winning state:
- objective, the biological imperative: survive and pass on your genes
- subjective, your own experience: live a life that feels fulfilling
- both (probably the most fulfilling winning state for a human being)

Here's the minimum loop to advance towards your chosen winning state:
1. assess where you stand on your lifeboard
2. plan what you think is your best next move
3. execute it
4. learn something new about the game

I made Claude talk to Grok @bot via Github.
I can now use Grok bot from within Claude Code.
x.com/fedeandri/stat…