Inkling’s Two-Terabyte Paradox

Learn why Inkling can activate 41 billion parameters per token while its BF16 release still carries a roughly two-terabyte serving estimate—and how to judge what its software demos and benchmark scores actually prove. We begin with the physical release: 108 main BF16 weight shards and about 1.905 terabytes of inspected repository files. Following one token through six of 256 routed experts plus two shared experts reveals how sparse computation can coexist with a nearly trillion-parameter resident system. Then we reconstruct what the public training story does disclose: 45 trillion multimodal pretraining tokens, Muon plus Adam optimization, NVIDIA GB300 NVL72 systems, a synthetic supervised-fine-tuning bootstrap, and more than 30 million asynchronous reinforcement-learning rollouts. Just as importantly, we keep the missing data manifest, compute, schedules, reinforcement-learning algorithm, reward definitions, environment mixture, and variance visible rather than pretending the pipeline is a reproducible recipe. The software demonstrations are traced as complete systems. Inkling generates a resume application before a separate browser-use agent fills it from saved profile data. The multiplayer game is the endpoint of a demanding specification, tools, tests, and 40 rounds of GPT Codex reviewer feedback—not a demonstrated one-shot build. The same attribution rule then follows the benchmark results back to their effort settings, harnesses, task versions, tests, and known integrity limits. By the end, you can repair the shortcut “a 41B model built a game”: Inkling is officially 975B total with 41B active per token, while the automatic repository count of about 952.38B remains unreconciled; and the published game belongs to an Inkling-centered review-and-revision system. The final takeaway is a reusable evidence threshold: a selected trace can show what is possible, conditioned task sets can show what was measured, and deployment readiness must be earned in the system you will actually run. Evidence in the video was inspected on July 16, 2026, one day after Inkling’s release, so the verdict is deliberately provisional. Chapters 0:00 Two Terabytes Before 41B 1:07 One Token, Six Routes 2:40 The Count That Will Not Close 3:44 Three Inputs, One Decoder 5:02 A Pipeline Is Not a Recipe 6:21 Effort Was Trained 7:57 The App Has Two Actors 9:15 The Game Has Forty Loops 10:49 The System Is the Demo 12:02 A Profile, Not a Rank 13:22 The Yardstick Moves 14:59 Repair the Headline 16:18 What the Release Earns