Ternary Bonsai 2 9B
A ternary-quantized 9B-parameter language model released by PrismML in GGUF format, with a file size significantly smaller than the 27B version in the same series.
Timeline (last 90 days)
PrismML released Ternary Bonsai 2 9B, whose GGUF file is only 2.25 GiB, while the 27B version is 6.70 GiB.
In laptop benchmarks, Ternary Bonsai 2 9B is 2.4–3.6 times faster than 27B.
In replay testing, the 9B downgraded 15 of 65 escalation requests already confirmed by the second model to clarify.
On task B, 43% of the 9B's answers returned 'intervention needed: true' together with 'continue as normal', violating the application-layer rules.
In task A, after reweighting by production volume, the 9B's allow proportion dropped from the 27B's 86.4% to 49.5%, clarify rose from 13.4% to 50.1%, and the weighted agreement rate was 61.2%.
In task B, after reweighting, about 17% of checks would be routed to the managed fallback due to verification failure, actually increasing the load that was originally intended to be reduced.
In task C, the 9B stayed consistent on 91 of the 103 cases the 27B marked as flagged or urgent, an agreement rate of 88.3%; but there were only five none cases, making it impossible to reliably estimate the false positive rate.
All Facts (7)
On task B, 43% of the 9B's answers returned 'intervention needed: true' together with 'continue as normal', violating the application-layer rules.
50%UnverifiedIn laptop benchmarks, Ternary Bonsai 2 9B is 2.4–3.6 times faster than 27B.
50%UnverifiedIn replay testing, the 9B downgraded 15 of 65 escalation requests already confirmed by the second model to clarify.
50%UnverifiedIn task C, the 9B stayed consistent on 91 of the 103 cases the 27B marked as flagged or urgent, an agreement rate of 88.3%; but there were only five none cases, making it impossible to reliably estimate the false positive rate.
50%UnverifiedIn task A, after reweighting by production volume, the 9B's allow proportion dropped from the 27B's 86.4% to 49.5%, clarify rose from 13.4% to 50.1%, and the weighted agreement rate was 61.2%.
50%UnverifiedIn task B, after reweighting, about 17% of checks would be routed to the managed fallback due to verification failure, actually increasing the load that was originally intended to be reduced.
50%UnverifiedPrismML released Ternary Bonsai 2 9B, whose GGUF file is only 2.25 GiB, while the 27B version is 6.70 GiB.
50%