27.36B total: 24.35B language backbone + 2.54B embeddings/LM head + 0.46B vision tower
262,144 tokens
Apache 2.0
Second-generation open-weight, aggressively compressed vision-language model from Prism ML, derived from Qwen3.8-27B with the base architecture unchanged. Ternary Bonsai 2 27B retains the base model's 64-block hybrid-attention language backbone, comprising approximately 75% linear-attention and 25% full-attention layers with SwiGLU feed-forward networks, RoPE, RMSNorm, native reasoning, text-and-image input, structured tool calling, and a 262,144-token context window. Prism ML's detailed model card counts 27.36B parameters, comprising a 24.35B language backbone, 2.54B embedding and language-model-head parameters, and an optional 0.46B, 27-block vision tower; the launch press release rounds the model to 27.8B. Language-model weights are represented end to end as ternary {-1, 0, +1} values with one shared FP16 scale per group of 128 weights, covering embeddings, attention projections, MLP projections, and the LM head. A fixed blockwise Hadamard rotation with block size 1,024 is folded into the stored weights, while the corresponding transform is applied to activations at runtime. Only 26.2M language-model parameters, approximately 0.098%, remain in higher precision. The compact PTQ1_0 GGUF densely packs ternary values at 1.75 effective bits per weight into a 5.95 GB language-model file, while the PQ2_0 version uses two-bit slots and occupies 7.21 GB in exchange for less expensive unpacking and substantially faster prompt processing on some hardware. Optional vision projectors add approximately 629 MB in Q8_0 or 931 MB in BF16. An MLX companion package for Apple Silicon occupies approximately 8.6 GB including the language model and vision tower. Prism ML reports an overall score of 83.9 across a 20-benchmark thinking-mode suite, retaining 98.2% of the full-precision Qwen3.8-27B baseline's 85.4. Reported category scores include 96.57 for math, 81.58 for coding, 82.66 for instruction following, 77.57 for agentic and tool-calling tasks, 83.95 for knowledge and reasoning, and 78.59 for vision. A narrower 14-benchmark table in the current model card reports an 84.78 average versus 86.32 for FP16, including 95.83 on AIME 2026, 90.07 on LiveCodeBench, 74.92 on BFCL v3, 75.49 on MMMU-Pro, and 74.00 on IFBench. Prism ML reports throughput of up to 143 tokens per second on an RTX 5090 and 46.8 tokens per second on an Apple M5 Max, with custom CUDA and Metal low-bit kernels. The packed GGUF formats require Prism ML's modified llama.cpp runtime because stock llama.cpp does not yet implement the required ternary hybrid-attention kernels and rotated-weight metadata; the MLX release likewise uses Prism ML's custom MLX stack. Official integrations and examples cover OpenAI-compatible serving, Pi, Hermes Agent, OpenClaw, Cline, multimodal workflows, MCP tools, and per-request reasoning budgets.