PrismML announced Ternary Bonsai 2 27B, an open-weight model release built around a 27-billion-parameter base and designed to reduce the storage and memory cost of local inference. The accompanying model repository makes the weights and packaging traceable for developers.
Compression is the product
PrismML says the release reaches 1.76 effective bits per parameter, fits in about 5.9 GB and supports a 262K context window. It is published under Apache 2.0 according to the announcement and model card, positioning the release for local or constrained deployments rather than only hosted inference.
A smaller footprint can change who can run a model
The model card records the released artifact and its stated runtime specifications. A model that fits within a smaller memory budget may broaden the hardware options for experimentation, but the practical speed, quality and compatibility depend on the runtime, quantization path and device.
Benchmark claims remain vendor-reported
PrismML reports 98.2% retention against its comparison baseline and additional benchmark results. The model repository corroborates the artifact and core specifications, but no independent reproduction of the benchmark suite was located for this edition, so the reported quality should not be treated as an independently verified score.