PrismML: Ternary Bonsai 2 27B

PrismML: Ternary Bonsai 2 27B is a 27B-parameter reasoning model from PrismML derived from Qwen3.8-27B, with a 262,144-token context window, text-only modality, and function-calling support. According to LLMBase, it supports coding, mathematics, tool calling, and image understanding, and applies ternary compression to The LLMBase profile attributes the model to provider "Prism Ml" and notes that PrismML: Ternary Bonsai 2 27B is not currently available in LLMBase Chat or the LLMBase Inference API, with benchmarks, pricing, and model details retained for research and comparison. The page preserves creator attribution to PrismML on a Q

The GenAI Secret Sauce weekly digest for September 12-18, 2026 explicitly cites PrismML's Ternary Bonsai 2 27B as a 27B model built on Qwen3.8 27B quantized to 1.76 bits per weight, yielding a 5.9 GB model size that retains 98.2% of its scores. This frames the release within a broader roundup that also covers IBM AppWo The digest places Ternary Bonsai 2 27B alongside other notable technical items of the week, including Anthropic's outside-evaluator plan, OpenAI's six misalignment incident reports, and a poll showing 63% of Americans see at least moderate risk in AI destroying humanity. Creator attribution to PrismML on a Qwen3.8 27B

BenchLM tracks Ternary Bonsai 2 27B with 21 published benchmark rows across 230 ranked models, reporting a capability score of 50.8/100 against a field median of 56.3. Category-level verified evidence includes Agentic (rank 59 of 154, percentile 62nd, 3/3 benchmarks), Coding (rank 64 of 154, percentile 59th, 4/4), Know Strengths highlighted by BenchLM include agentic performance at rank 59, described as particularly useful for coding agents, browser research, and computer-use workflows. Coverage caveats noted on the page are that Reasoning and Multilingual categories have zero published benchmarks and that speed/time-to-first-token a