The card lists a total file size of 129,565,620,627 bytes, sharded safetensors, and a region tag of us. The weights are unquantized.

The 80.52 TerminalBench 2.1 score, the $32 million training cost, and the 12-trillion-token corpus all originate on Ricursive's product page, not in the artifact where the weights live.

A 64.8B dense Transformer in BF16 is a deployable object that runs on DGX Spark-class hardware without quantization. The vendor's page describes the architecture as optimized for Nvidia hardware.

What the card does not prove is that the model scores 80.52, or that training cost $32 million, or that it saw 12 trillion tokens. The training recipe sits on a marketing page, not in a reproducible config file.

No page reports GPU hours, FLOPs consumed, the specific pretraining GPU models, or inference tokens-per-second on DGX Spark. The 1.4M rollout figure and the 70x comparison to Reflection Beam are vendor-stated, not independently measured.

The buyer inherits a deployable weight file whose headline number lives one page away from the weights themselves.

The draft restated the same gap three times in slightly different words. This version leads with the absence, states the card contents once, then walks the vendor-page claims and the missing telemetry in two tight blocks before closing on the buyer's position. Rhythm shifts from one-sentence to two-sentence paragraphs to avoid the flat cadence of the original.