1.88B / 4.21B / 8.95B language model + 0.33B / 0.33B / 0.46B vision tower
262,144 tokens
Apache 2.0
Grouped release of three compact second-generation ternary vision-language models derived from Qwen3.5-2B, Qwen3.5-4B, and Qwen3.5-9B with their base architectures unchanged. The 2B model contains approximately 1.88B language-model parameters across 24 blocks plus a 0.33B, 24-block vision tower; the 4B contains approximately 4.21B language-model parameters across 32 blocks plus the same 0.33B vision tower; and the 9B contains approximately 8.95B language-model parameters across 32 blocks plus a 0.46B, 27-block vision tower. All three retain Qwen3.5's 262,144-token context window, native reasoning and non-thinking modes, image understanding, and structured tool-calling capabilities. Their hybrid-attention language backbones use a repeating 3:1 ratio of linear-attention to full-attention layers with SwiGLU MLPs, RoPE, and RMSNorm. Language-model matrix weights are represented as ternary {-1, 0, +1} values with FP16 group-wise scales and a group size of 128. A fixed blockwise Walsh-Hadamard rotation is folded into the stored weights, while normalization weights and the linear-attention state path remain in F32. The smallest PTQ1_0 GGUF language-model packs occupy approximately 0.43 GB, 0.95 GB, and 2.00 GB, versus 3.76 GB, 8.41 GB, and 17.91 GB for the corresponding BF16 language models, making them approximately 8.8-9.0 times smaller. Prism ML also provides faster-to-unpack PQ2_0 packs, Q2_0 packs compatible with recent upstream llama.cpp, MLX two-bit packages for Apple Silicon, Q8_0 and BF16 vision projectors, and optional MTP and DFlash2 speculative-decoding components for the 4B and 9B variants. Across Prism ML's evaluation suite, reported overall averages are 61.81 versus 63.95 for the Qwen3.5-2B baseline, 73.36 versus 77.53 at 4B, and 77.02 versus 80.96 at 9B. Relative to their same-size BF16 baselines, the ternary models retain 98.2-107.8% of math performance, 90.0-95.5% of agentic and tool-calling performance, 100.1-115.0% of coding performance, 100.0-121.5% of instruction-following performance, and 89.0-92.5% of vision performance. The 9B reports 82.5 on AIME 2026, 89.02 on HumanEval+, 71.81 on BFCL v3, and 78.02 on BFCL v4 Mobile. Reported batch-one decode throughput using PTQ1_0 reaches approximately 603.6, 368.6, and 248.6 tokens per second for the 2B, 4B, and 9B models on an RTX 4090. MLX throughput on an Apple M5 Pro reaches approximately 276.3, 145.2, and 91.7 tokens per second, while an iPhone 17 Pro Max reaches approximately 89.9, 39.8, and 22.3 tokens per second. Prism ML reports 3.6-4.2 times lower generation energy per token than BF16 on an RTX 4090 and 1.7-2.0 times lower energy on an M5 Pro. The models were trained using Google TPU v5 hardware. Deployment uses Prism ML's custom llama.cpp kernels for the PTQ1_0 and PQ2_0 formats or its MLX-compatible runtime, with official examples covering OpenAI-compatible serving, multimodal prompting, local tool use, and speculative decoding.