PrismML launches ternary Bonsai models

This is a strong compression story: PrismML is no longer just chasing the smallest possible model, it’s optimizing for the more useful point where a little extra memory buys a lot more capability.

• –The core design is fully ternary, with weights constrained to `{-1, 0, +1}` across embeddings, attention, MLPs, and the LM head.
• –PrismML claims the 8B model scores 75.5 average benchmark points, about 5 points better than its 1-bit 8B predecessor, while staying at 1.75 GB.
• –The deployment angle is the real hook: native MLX support on Apple devices and reported throughput of 82 toks/sec on M4 Pro make this feel practical, not just academic.
• –Apache 2.0 licensing matters here, because it lowers friction for experimentation and downstream packaging.
• –The big question is how these numbers hold up outside PrismML’s own benchmark setup, especially across real workloads and longer-context use.