Loading PrismML's 5.95 GB Bonsai 2 requires the company's own llama.cpp fork

PrismML says the ternary weights preserve 98.2 percent of the full-precision model's benchmark performance while cutting the footprint by roughly nine times ^[2]. Divide the launch announcement's two averages and you get 98.24 percent ^[18]. Divide the model card's and you get 98.22 percent ^[19]. Different suites, different benchmark counts, same ratio to one decimal place, so the disagreement in the underlying scores never reaches the headline number. PrismML does not reconcile the two sets of figures in its published materials ^[15].

The compression claim holds on its own terms. PrismML puts the ideal ternary language-model footprint at about 5.8 GB, which is 9.3 times smaller than the FP16 baseline, and the file you actually download is 9.1 times smaller ^[4][20]. Both figures cover language weights. Image input needs the optional 8-bit vision component, another 0.63 GB, taking the compact build to 6.58 GB; the alternative packing, designed to reduce unpacking work, is 7.21 GB, or 7.84 GB with vision ^[5][21].

The reason stock llama.cpp cannot open these files is that it does not implement PrismML's custom ternary formats or its Hadamard activation transform ^[11]. PrismML says its backends cover CUDA, Metal and CPU execution ^[12], and there is an MLX companion for Apple Silicon ^[6]. Apache 2.0 on the weights does not remove the integration work: until the formats reach widely used upstream inference software, developers have to adopt PrismML's kernels and runtime build ^[13]. The download size does not include that cost.

The aggregate average hides where the quality went. Math is nearly flat, 96.57 against 97.06 ^[8]. Coding comes out slightly above the FP16 baseline, 89.42 to 89.07 ^[8]; I would read a compressed model beating its own baseline as the coding suite running out of resolution at that level. Knowledge and reasoning lost 5.69 points, a 6.7 percent relative drop ^[22]. Vision lost 5.17 points ^[23]. On OCR Bench v2 the compressed model scored 56.88 against 60.99 for FP16, and the model card places it below a conventional 2-bit build as well ^[10].

For the announcement's 143 tokens per second to transfer, you need an RTX 5090, PrismML's runtime, the same packing and single-stream decode ^[14]. The model card's own measurement on the same GPU is 129.9 tokens per second for one packing and 120.5 for the smaller one, about 10 percent under the announcement ^[7][24]. The documents also describe different models: 27.8 billion parameters in the announcement, 27.36 billion on the card ^[14][7].

PrismML says it emerged from Caltech research, with backing or support from Khosla Ventures, Cerberus, Caltech and Google, and continuing support from Samsung ^[16]. Dealroom reported in April that the company had raised $16.25 million through a SAFE and seed financing from Khosla, Cerberus and Caltech, and the round's valuation and ownership terms were not reported ^[17]. For a product where retrieval supplies the facts and the model writes code, 5.95 GB plus a pinned fork is a defensible trade in my view. Where the product leans on the model's own knowledge, the 5.69-point gap is the figure I would try to reproduce on my own evaluation set before building the runtime.