greenpolo/langslice-gemma-4-E4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 17, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

LangSlice-Gemma-4-E4B is a 7.9 billion parameter Gemma 4 E4B-IT fine-tune developed by greenpolo, specifically designed for estimating the anterior-posterior (AP) coordinate of histological mouse brain sections. This model excels at precisely localizing brain slices within a coronal atlas, achieving a mean absolute error of 1.228 mm (9.32% of the AP axis) on the SliceBench dataset. It is optimized for single-slice mouse coronal AP estimation, offering a specialized solution for neuroscientific research.

Loading preview...

Model Overview

LangSlice-Gemma-4-E4B v1.0 is a specialized fine-tuned version of the Gemma 4 E4B-IT model, developed by greenpolo. This 7.9 billion parameter model is engineered to accurately estimate the anterior-posterior (AP) coordinate of histological brain sections, specifically for mice, within a matching coronal atlas. It has been trained on a diverse dataset including adult (Allen mouse 25 µm) and developmental (DeMBA P9–P35, ADMBA P4–P56) coronal sections.

Key Capabilities & Performance

  • Precise AP Coordinate Estimation: The model's primary function is to localize single-slice mouse coronal brain sections.
  • High Accuracy: On the SliceBench small coronal mouse subset (n = 112), LangSlice-Gemma-4-E4B v1.0 achieves a mean absolute error of 1.228 mm, which translates to 9.32% of the 13.2 mm AP axis. This represents a significant improvement over the base unsloth/gemma-4-E4B-it model, which had a 18.02% error.
  • Specialized Scope: The model is intentionally scoped for single-slice mouse coronal AP estimation. Performance for sagittal, horizontal, or rat sections is noted to be significantly weaker.

Usage & Integration

  • Serving: Can be served via vLLM with specific configurations for max-model-len, enable-auto-tool-choice, and tool-call-parser.
  • Local Inference: GGUF quantized versions (Q4_K_M, Q8_0, BF16) are available for use with llama.cpp's llama-mtmd-cli, requiring pairing with the bundled mmproj-BF16.gguf for image input.
  • Chat Template: Inference must utilize the bundled chat template.

License

The model weights are governed by the Gemma 4 Terms of Use.

Finetunes
1 models