PrismML adds three smaller models to its Ternary Bonsai 2 lineup
The AMW Read
The release incrementally extends PrismML's ternary model lineup to smaller devices, with impact concentrated in local inference and company-reported performance still requiring workload-specific validation.
PrismML adds three smaller models to its Ternary Bonsai 2 lineup
PrismML has released Qwen3.5-based Ternary Bonsai 2 models with 2B, 4B and 9B parameters, extending a lineup that already includes a 27B version. Their ternary weights encode parameters as {-1, 0, +1}, with reported storage footprints of approximately 0.43GB, 0.95GB and 2GB, respectively—more than nine times smaller than full-precision versions. The weights are available under Apache 2.0, with deployment through llama.cpp on NVIDIA GPUs or MLX on Macs, iPhones and iPads.
The release broadens the foundation-model market's options for local deployment by pairing smaller weight files with reported improvements in reasoning, visual understanding, tool calling and instruction following. PrismML says throughput reaches up to five times that of full-precision versions on hardware including the RTX 4090 and M5 Pro, with threefold energy efficiency. It also reports up to 34 tokens per second for the 9B model and 120 for the 2B model on an iPhone 18 Pro Max. These are company-reported results; the article provides no detailed test methodology to establish how broadly they apply.
For builders evaluating local agents, the concrete opportunity is to test whether these smaller models can complete their workflows within device constraints. PrismML claims the 9B model consumes about 40% fewer tokens on standard instruction and everyday math tasks, but that does not establish equivalent savings across agent workloads. A useful evaluation would compare task completion, tool-call accuracy, latency and total memory use on the target device: compact weight storage alone does not establish application reliability.
