Ternary Bonsai 2 27B: benchmark results
PrismML released ternary derivative of Qwen3.8-27B under Apache2.0; distinct from Qwen and prior Bonsai generation. Effort/harness suffixes are preserved as exact evaluation configurations. Current released weight revision is not asserted to be the unpublished exact evaluation revision. Approximate27.36B multimodal total explicitly stated by technical report Table1 and official model card; Hub GGUF total26895998464 counts language tensors only. Launch prose27.8B is superseded by the consistent technical specification, not used as exact tensor count. Provider: PrismML. Released 2026-09-17. Access: Open.
Aggregate Rating 1702 ± 28, rank #338 of 10916 rated models, from 48 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| PrismML Bonsai 2 27B - Medium - IFBench | 55.33 | Prompt-level strict accuracy (%) | 100 |
| PrismML Bonsai 2 27B - XHigh - AIME26 | 95.83 | Mean accuracy avg@8 (%) | 100 |
| PrismML Bonsai 2 27B - XHigh - IFBench | 74 | Prompt-level strict accuracy (%) | 100 |
| PrismML Bonsai 2 27B - XHigh - LiveCodeBench v6 | 90.07 | Pass@1 (%) | 100 |
| ReasonScape R12 | 901.1 | ReasonScore | 78.7 |
| PrismML Bonsai 2 27B - XHigh - AA-LCR | 77 | Accuracy (%) | 66.7 |
| PrismML Bonsai 2 27B - XHigh - AIME25 | 95 | Mean accuracy avg@8 (%) | 66.7 |
| PrismML Bonsai 2 27B - XHigh - GSM8K | 96.66 | Numeric-equivalence accuracy (%) | 66.7 |
| PrismML Bonsai 2 27B - XHigh - HumanEval+ | 95.12 | Pass@1 (%) | 66.7 |
| PrismML Bonsai 2 27B - XHigh - IFEval | 91.31 | Prompt-level strict accuracy (%) | 66.7 |
| PrismML Bonsai 2 27B - XHigh - OmniDocBench v1.6 | 89.13 | Overall document-parsing composite (%) | 66.7 |
| BenchLM | 53.8 | Overall Score | 60.6 |
Interactive version: theaggregate.ai/model?slug=ternary-bonsai-2-27b · How It Works · Data refreshed daily, snapshot 2026-11-08.