News, Fast

Atom Brief

Jul 15, 2026 · Archive
TechUPDATED

Hugging Face Details What It Takes To Actually Run Inkling

Thinking Machines' 975-billion-parameter Inkling now runs on five frameworks day-one, including a Transformers update shipped the same day.

  • The full BF16 checkpoint needs 2 terabytes of memory; the NVFP4 version needs 600 gigabytes.
  • Unsloth's 1-bit GGUF build cuts that memory requirement by 95 percent for local machines.
  • Hugging Face offered free inference through its Inference Providers for the first two hours after launch.

Why it matters: Access to a free model is rationed by the two terabytes of memory required to run it.

Hugging Face blog (Thinking Machines) — "Welcome Inkling by Thinking Machines" ↗ · Jul 15, 20267/15/26 · Updated Jul 15