What's Changed

• Ollama's app now supports searching when running DeepSeek-V3.1, Qwen3 and other models that support tool calling.
• Flash attention is now enabled by default for Gemma 3, improving performance and memory utilization
• Fixed issue where Ollama would hang while generating responses
• Fixed issue where qwen3-coder would act in raw mode when using /api/generate or ollama run qwen3-coder <prompt>
• Fixed qwen3-embedding providing invalid results
• Ollama will now evict models correctly when num_gpu is set
• Fixed issue where tool_index with a value of 0 would not be sent to the model

Experimental Vulkan Support

Experimental support for Vulkan is now available when you build locally from source. This will enable additional GPUs from AMD, and Intel which are not currently supported by Ollama. To build locally, install the Vulkan SDK and set VULKAN_SDK in your environment, then follow the developer instructions. In a future release, Vulkan support will be included in the binary release as well. Please file issues if you run into any problems.

New Contributors

• @yajianggroup made their first contribution in #12377
• @inforithmics made their first contribution in #11835
• @sbhavani made their first contribution in #12619

Full Changelog: v0.12.5...v0.12.6

What's Changed

• Thinking models now support structured outputs when using the /api/chat API
• Ollama's app will now wait until Ollama is running to allow for a conversation to be started
• Fixed issue where "think": false would show an error instead of being silently ignored
• Fixed deepseek-r1 output issues
• macOS 12 Monterey and macOS 13 Ventura are no longer supported
• AMD gfx900 and gfx906 (MI50, MI60, etc) GPUs are no longer supported via ROCm. We're working to support these GPUs via Vulkan in a future release.

New Contributors

• @shengxinjing made their first contribution in #12415

Full Changelog: v0.12.4...v0.12.5-rc0

What's Changed

• Flash attention is now enabled by default for Qwen 3 and Qwen 3 Coder
• Fixed minor memory estimation issues when scheduling models on NVIDIA GPUs
• Fixed an issue where keep_alive in the API would accept different values for the /api/chat and /api/generate endpoints
• Fixed tool calling rendering with qwen3-coder
• More reliable and accurate VRAM detection
• OLLAMA_FLASH_ATTENTION can now be overridden to 0 for models that have flash attention enabled by default
• macOS 12 Monterey and macOS 13 Ventura are no longer supported
• Fixed crash where templates were not correctly defined
• Fix memory calculations on NVIDIA iGPUs
• AMD gfx900 and gfx906 (MI50, MI60, etc) GPUs are no longer supported via ROCm. We're working to support these GPUs via Vulkan in a future release.

New Contributors

• @Fachep made their first contribution in #12412

Full Changelog: v0.12.3...v0.12.4-rc3

New models

• DeepSeek-V3.1-Terminus: DeepSeek-V3.1-Terminus is a hybrid model that supports both thinking mode and non-thinking mode. It delivers more stable & reliable outputs across benchmarks compared to the previous version:

  Run on Ollama's cloud:

  ollama run deepseek-v3.1:671b-cloud
  

  Run locally (requires 500GB+ of VRAM)

  ollama run deepseek-v3.1
  

• Kimi-K2-Instruct-0905: Kimi K2-Instruct-0905 is the latest, most capable version of Kimi K2. It is a state-of-the-art mixture-of-experts (MoE) language model, featuring 32 billion activated parameters and a total of 1 trillion parameters.

  ollama run kimi-k2:1t-cloud
  

What's Changed

• Fixed issue where tool calls provided as stringified JSON would not be parsed correctly
• ollama push will now provide a URL to follow to sign in
• Fixed issues where qwen3-coder would output unicode characters incorrectly
• Fix issue where loading a model with /load would crash

New Contributors

• @gr4ceG made their first contribution in #12385

Full Changelog: v0.12.2...v0.12.3

Web search

A new web search API is now available in Ollama. Ollama provides a generous free tier of web searches for individuals to use, and higher rate limits are available via Ollama’s cloud. This web search capability can augment models with the latest information from the web to reduce hallucinations and improve accuracy.

What's Changed

• Models with Qwen3's architecture including MoE now run in Ollama's new engine
• Fixed issue where built-in tools for gpt-oss were not being rendered correctly
• Support multi-regex pretokenizers in Ollama's new engine
• Ollama's new engine can now load tensors by matching a prefix or suffix

Full Changelog: v0.12.1...v0.12.2

New models

• Qwen3 Embedding: state of the art open embedding model by the Qwen team

What's Changed

• Qwen3-Coder now supports tool calling
• Ollama's app will now longer show "connection lost" in error when connecting to cloud models
• Fixed issue where Gemma3 QAT models would not output correct tokens
• Fix issue where & characters in Qwen3-Coder would not be parsed correctly when function calling
• Fixed issues where ollama signin would not work properly on Linux

Full Changelog: v0.12.0...v0.12.1

Cloud models

Cloud models are now available in preview, allowing you to run a group of larger models with fast, datacenter-grade hardware.

To run a cloud model, use:

ollama run qwen3-coder:480b-cloud

• View all cloud models
• Blog post

What's Changed

• Models with the Bert architecture now run on Ollama's engine
• Models with the Qwen 3 architecture now run on Ollama's engine
• Fix issue where older NVIDIA GPUs would not be detected if newer drivers were installed
• Fixed issue where models would not be imported correctly with ollama create
• Ollama will skip parsing the initial <think> if provided in the prompt for /api/generate by @rick-github

New Contributors

• @egyptianbman made their first contribution in #12300
• @russcoss made their first contribution in #12280

Full Changelog: v0.11.11...v0.12.0