●Change8

v0.11.7

📦 ollama
✨ 4 features🐛 4 fixes🔧 2 symbols

Summary

This release introduces the DeepSeek-V3.1 model and the preview of Turbo mode for running large models. Several bugs related to model loading, thinking tags, and tool call parsing have also been resolved.

Migration Steps

  1. To use Turbo mode via the CLI, create an account on ollama.com/signup and follow the documentation to authenticate your Ollama installation.
  2. When running DeepSeek-V3.1 via CLI with Turbo enabled, use: OLLAMA_HOST=ollama.com ollama run deepseek-v3.1

✨ New Features

  • DeepSeek-V3.1 model is now available via Ollama.
  • DeepSeek-V3.1 supports hybrid thinking, configurable via the 'think' parameter in the Ollama API.
  • Thinking can be enabled/disabled in the Ollama CLI using /set think or /set nothink commands.
  • Introduction of Turbo mode (in preview) to run large models like DeepSeek-V3.1 (671B parameters) using cloud hardware.

🐛 Bug Fixes

  • Fixed an issue preventing multiple models from loading on CPU-only systems.
  • Ollama now correctly handles models that skip outputting the initial <think> tag (e.g., DeepSeek-V3.1).
  • Fixed an issue where text was emitted when a model did not output an opening <think> tag.
  • Fixed incorrect parsing of tool calls that contained '{' or '}' characters.

🔧 Affected Symbols

DeepSeek-V3.1<think>