PrismML / prismml/ternary-bonsai-2-27b

Ternary Bonsai 2 27B - access through LLMTR

Ternary Bonsai 2 27B is PrismML's compressed build of Qwen3.8 27B: its weights take only the values -1, 0 and +1 (about 1.76 effective bits per weight), making it more than nine times smaller than the full-precision original. It offers a 262,144-token context window, text and image input with text output, native tool calling including forced tool_choice and parallel calls, JSON object mode and JSON Schema structured output. Step-by-step reasoning is ON by default; the reasoning trace is returned separately from the answer, reasoning tokens are billed as output, and reasoning can be switched off per request. Graded reasoning_effort levels are not supported, only on or off. The context window is SHARED between prompt and output: there is no separate output ceiling beneath it, and the largest output you can request shrinks as the prompt grows. No prompt-cache discount is applied; every request bills at the full input rate.