●Change8

v0.13.0

📦 ollamaView on GitHub →
✨ 7 features🐛 2 fixes🔧 7 symbols

Summary

This release introduces support for DeepSeek-OCR, Cogito-V2.1, and DeepSeek-V3.1 architecture, alongside a new performance benchmarking tool and significant engine optimizations for KV caching and GPU detection.

Migration Steps

  1. To use the new bench tool, install Go and run 'go run ./cmd/bench -model <model_name>' from the root of the Ollama repository.

✨ New Features

  • Added support for DeepSeek-OCR model for high-precision document processing and optical 2D mapping.
  • Added support for Cogito-V2.1 instruction-tuned generative models.
  • Added support for DeepSeek-V3.1 architecture in the Ollama engine.
  • Introduced a new 'bench' tool in the GitHub repository for testing model performance.
  • Improved KV cache performance to eliminate the need for defragmentation.
  • Improved CPU and memory detection within containers and cgroups.
  • Improved VRAM information detection for AMD GPUs.

🐛 Bug Fixes

  • Fixed performance regressions on CUDA introduced in version 0.12.11.
  • Fixed missing Vulkan libraries in Linux installation packages.

🔧 Affected Symbols

DeepSeek-OCRCogito-V2.1DeepSeek-V3.1cmd/benchCUDAVulkanKV cache