Linnaeus-0.1.0-2B — MLX 8bit (Apple Silicon)

MLX-converted Linnaeus decision model for macOS and iOS (via mlx-swift-lm, which already ships Qwen35.swift). Text decisions only — the vision tower is dropped by the MLX weight loader.

Decision contract

Score = logits[marker_pos, score_row_id] where markers are <|fim_suffix|> (one appended after each candidate option). Contract in linnaeus-runtime.json.

from linnaeus.mlx_predictor import MlxPredictor
p = MlxPredictor("Linnaeus-0.1.0-2B-MLX-8bit")
out = p.predict(state, {"q": {"type": "choice", "instructions": "...",
                              "criteria": {...}}})

Measured on JevBench v1.2.2 (231 tasks, M4): 8bit: 70.56% (1.9GB). Upstream reference 65.8%; CUDA fp32 pipeline 73.16%.

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.

Downloads last month: 14

Safetensors

Model size

2B params

Tensor type

U32

BF16

F32

MLX

Hardware compatibility

8-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for pi-dal/Linnaeus-0.1.0-2B-MLX-8bit

Base model

Qwen/Qwen3.5-2B-Base

Finetuned

Qwen/Qwen3.5-2B

Adapter

pi-dal/Linnaeus-0.1.0-2B

Quantized

(4)

this model