pi-dal/Linnaeus-0.1.0-2B-MLX-VLM-4bit

Linnaeus-0.1.0-2B — MLX VLM 4bit (Apple Silicon, multimodal)

Multimodal MLX build of the Linnaeus decision model: text + image decisions on macOS and iOS (via mlx-swift-lm Qwen35.swift + vision tower). Language stack quantized to 4bit; vision tower kept bf16.

Decision contract

Score = logits[marker_pos, score_row_id] at each <|fim_suffix|> marker. Contract in linnaeus-runtime.json. Image questions prepend <|vision_start|><|image_pad|><|vision_end|> and pass pixel_values through the mlx-vlm processor.

Measured on JevBench v1.2.2 text tasks: ~67% text + image-capable (1.6GB, vision tower bf16). Image path verified end-to-end against torch (prob deltas < 0.001 on synthetic probes). Upstream 65.8%.

Predict path note

Multi-question predict() reuses a shared prefix: the state (+ image tokens) is forwarded once and each question forwards only its suffix — measured 1.4x on 8 text questions, 4.5x on 6 image questions. Internal to MlxPredictor; no contract change.