Ternzip: lossless pack of Ternary Bonsai 8B

ternzip-8b.trits is a byte-exact, third-party recompression of Prism ML’s Ternary Bonsai 8B Q2_0 GGUF. The complete archive is 1,623,736,952 bytes (1.624 GB), including the weights, FP16 scales, tokenizer and metadata. It saves 558.45 MB (25.59%) against Prism’s 2,182,184,672-byte file. Archive download · Archive checksum · Original GGUF

How it works

Ternzip matches whole rows and columns across selected layers, even when their indices differ or their signs flip. Those matches help a fast entropy coder called rANS encode the actual weights using probabilities fitted to this checkpoint. Most tensors store the indices of their FP16 scales as exact differences from row and column predictions. Decoding recovers every weight and scale bit. Worked example and encoding details

Restore or run

Download the archive and the Ternzip v0.1.0 source. From the source directory, run the following command with CPython 3.14 and stable Rust installed. It writes Ternary-Bonsai-8B-Q2_0.gguf and checks Prism’s published SHA-256. Versioned GitHub loader and instructions

./ternzip restore /path/to/ternzip-8b.trits

Direct use of the archive requires the Ternzip custom loader. On an Apple M4 Pro, load time was about 19 seconds with dependencies ready. Ten greedy story and code runs had a median of 61.38 output tokens/s after prompt processing. Model-related allocations including reserve were 1,999,704,327 bytes. Whole-process physical memory after generation was 2.99–3.35 GB. For stock llama.cpp, Minarut publishes a 2.12 GB TQ2_0 GGUF. Benchmark measurements · Minarut GGUF

Prior work

The related work covers ternary inference, five-trit packing and lossless block matching. TernaInfer · Intel BITCOS · llama.cpp TQ1_0 · Storer and Helfgott, 1997

License and attribution

Created using Bonsai by Prism ML. Optimatech Labs modified the distribution files by changing the checkpoint’s storage format. This is an unofficial, third-party release. Prism ML developed Ternary Bonsai 8B from Qwen3-8B, copyright 2024 Alibaba Cloud. Apache License 2.0 · Prism NOTICE · Qwen license and copyright

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for optimatechlabs/ternzip-8b

Quantized
(3)
this model

Paper for optimatechlabs/ternzip-8b