bitnet-kv260-weights

The weights of microsoft/bitnet-b1.58-2B-4T, repacked into the byte layout that the FPGA fabric of a Xilinx Kria KV260 streams directly.

They are used by bitnet-kv260, which runs all 30 layers of the model on a KV260: every ternary matrix-vector product is streamed through four engines in the programmable logic, while attention, the norms and the exact top-K rescore of the output head run on the board's four Cortex-A53 cores.

These files are a repacking, not a retraining. No weight value is changed. The setup script of that project downloads them, checks every file against the sizes and MD5 sums recorded in manifest.json and head_t.json, and refuses to continue on a mismatch.

Files

file	bytes	what it is
model3.bin	417,546,240	every ternary matrix of all 30 layers, base-3 packed, five trits a byte, 80 weights a 128-bit beat, row-major per neuron
embed_bf16.bin	656,670,720	the token embedding, bf16, unchanged from the checkpoint
head_i8.bin	328,335,360	the tied output head, quantised per row to int8
head3_t.bin	65,667,072	a ternary approximation of the output head, base-3 packed, streamed by the fabric to shortlist candidates
head_scale.bin	513,024	the per-row scales of head_i8.bin, float32
head_t_scale.bin	513,024	the per-row scales of head3_t.bin, float32
norms.bin	1,761,280	every norm weight, float32, unchanged from the checkpoint
manifest.json	78,011	the geometry, the offsets of all 210 matrices, and the size and MD5 of every file
head_t.json	1,099	the ternary head's layout, its rule, and its size and MD5

tokenizer.json is deliberately not here. The project fetches it from microsoft/bitnet-b1.58-2B-4T directly.

How the output head works

The head is two stages and the result is exactly the argmax of the int8 head. Stage one streams the ternary approximation through the fabric and takes the best 256 rows. Stage two recomputes those 256 rows exactly from head_i8.bin on the processors and takes the argmax. Measured over 2,115 real hidden states, the true argmax never fell below rank 56 in the approximation, so a shortlist of 256 covers it with a factor of 4.6 in hand.

Licence and attribution

MIT, because these files are derived from microsoft/bitnet-b1.58-2B-4T, which is MIT.

Copyright (c) Microsoft Corporation. MIT License.

embed_bf16.bin and norms.bin are byte-for-byte the checkpoint's own tensors. The rest are derivatives of its tensors. If you redistribute them, the MIT notice travels with them.

The original model is described in arXiv:2504.12285. Microsoft's own model card notes: "We do not recommend using BitNet b1.58 in commercial or real-world applications without further testing and development."

Note that the code of the KV260 project is Apache 2.0, for its patent grant over the hardware architecture. Only these weight files are MIT. The two licences are separate and both apply where they say they do.