The notebook, published verbatim.

These pages render the repository's docs directory as-is. The numbering is the repo's own; when a doc changes in the repo, it changes here.

• 01The .trit formatA .trit file is a single self-describing container holding one quantized model: an embedded copy of the source config.json, a tensor table, and a payload of ternary and dense tensors.
• 01aArchitectureA BitNet b1.58 linear layer computes y = W·x where every weight is -1, 0 or +1 and activations are int8. The inner loop is therefore not multiply-accumulate but select-accumulate:
• 02RoadmapThe original roadmap was six numbered weeks ending in an FPGA demo box, written when the plan was "prove the primitive in simulation, buy a board, ship the demo." That plan produced everything it promised through its own
• 03Research notes1. BinaryConnect / BNN (2015–2016) — binary weights work for small nets; accuracy collapse at scale. 2. TWN / TTQ (2016–2017) — ternary {-1, 0, +1} recovers much of the loss; the zero state matters (it's learned sparsity
• 03bNumericsSource: TRITSIMSTATS recorder over two prompts ("The capital of France is", the 18-token variant) with 16 greedy decode steps each, on the pinned BitNet b1.58 2B4T checkpoint. Values are the max over both runs; count is
• 04Benchmark methodEvery public performance claim traces to a script in tritbench/ and a raw-data CSV committed to the repo. If a number can't be reproduced from a fresh clone, it doesn't get posted.
• 04bBoard memoDecision requested: which FPGA board (if any) to buy for the first hardware bring-up. Written after Phase 2; numbers below are measured, not guessed. Following the ternoise doctrine: simulation-first, board chosen from s
• 05Positioningtritium runs real language models on $50, single-digit-watt hardware — fully offline — by building the first complete hardware-native stack for ternary (1.58-bit) LLMs.
• 06GatesThe frozen reference for the hardware-agnostic pivot. Every number here was measured on the pre-pivot tree at commit c3351d1 (v0 .trit, scalar kernel, single-threaded) before any pivot code landed.