Ollama

This release focuses on improving hardware acceleration compatibility, particularly for CUDA 6.x GPUs, and cleans up experimental features and dead code. It also addresses CUDA toolkit lookup issues.

This release focuses on improving hardware compatibility, particularly for CUDA and ROCm, and cleans up experimental code paths. Key changes include enabling FA on older CUDA GPUs and removing deprecated environment flags.

This release focuses on improving hardware compatibility, particularly for CUDA 6.x GPUs, and cleans up internal code, including the removal of an experimental server flag.

This release introduces significant performance improvements for Gemma 4 on Apple Silicon by leveraging multi-token prediction (MTP). It also includes updates to the underlying MLX and llama.cpp engines.

This release introduces an improvement to tool call detection logic and updates underlying dependencies, including llama.cpp and mlx.

This release introduces new auto-installation features for models like Claude Code and opencode, alongside numerous stability and performance improvements across GPU handling (Vulkan, CUDA presets) and model loading/generation.

This release introduces new auto-installation features for models like Claude Code and opencode, alongside numerous stability and performance improvements across Vulkan, MLX, and model loading mechanisms. Key fixes include correcting GPU classification on Windows and improving speculative decoding.

This release introduces new auto-installation features for models like Claude Code and opencode, alongside numerous performance and stability fixes across Vulkan, speculative decoding, and model offloading mechanisms. It also updates the underlying llama.cpp dependency.

This release introduces the new Cohere2MoE model and updates the underlying llama.cpp library to revision b9672.

This release introduces the new Cohere2MoE model and updates the underlying llama.cpp library to revision b9672.

This release introduces the new Cohere2MoE model and updates the underlying llama.cpp library to revision b9672.

This release introduces support for the Cohere2Moe architecture and resolves several bugs related to LFM2 parsing and token output limitations in agent use cases. Ollama now enforces context window limits on single messages.

This release introduces support for the Cohere2Moe architecture and resolves several bugs related to LFM2 parsing and token output limitations in agent use cases. Ollama now enforces context window limits on single messages.

This release updates the underlying llama.cpp library to commit b9637.

This release updates the underlying llama.cpp library to commit b9637.

This release includes improvements to MLX MTP caching, hardening of mlxrunner layers, and a fix for launch provider drift. Prompt caching logic has also been decoupled from context shifting.

This release introduces Hermes Desktop as a new native interface for the Hermes agent, accessible via `ollama launch hermes-desktop`, and aligns the OpenAI-compatible API model listing.

This release includes documentation updates, specifically regarding Zod examples, and fixes the native configuration path for Hermes on Windows.

This release includes documentation updates regarding cloud model retirement, an MLX backend enhancement for the embedding layer, and a new integration with Oh My Pi.

This release addresses a critical crash related to the gemma4:12b model and includes an integration fix for Hermes on Windows.

This release introduces documentation for Cline CLI integration and fixes an issue with Hermes installation on Windows, alongside an update to the underlying llama.cpp version.

This release includes an update to the underlying llama.cpp version and fixes an issue related to cleaning up the llama-server process on Windows.

This release updates the underlying llama.cpp version and includes a fix for properly terminating the llama-server process on Windows during cleanup.

This release introduces support for the new gemma4-12b model within the models module.

This release introduces support for Cline CLI and Qwen integration, alongside various stability improvements and fixes related to llama-server and model loading.

This release introduces support for Cline CLI and Qwen integration, alongside numerous stability improvements and fixes related to llama-server and model loading.

This release introduces new features like Cline CLI support and Qwen code integration, alongside several bug fixes related to model limits, server counts, and markdown handling. It also updates the underlying llama.cpp version.

The OpenAI Codex App is now available on Ollama, integrating local and cloud models for coding workflows. This release also introduces a built-in browser and code review mode, alongside MLX sampler improvements for Apple Silicon.

This release introduces memory trace logging for MLX and integrates Codex application support into the launch mechanism.

This release introduces memory trace logging for MLX and integrates Codex application support via the launch mechanism.

This release introduces support for vision models with image inputs when launching opencode via ollama and fixes an issue with Claude tool result formatting for local image paths.

This release introduces support for vision models with image inputs when launching opencode via ollama and fixes an issue with Claude tool result formatting for local image paths.

This pre-release updates the architecture to directly support llama.cpp, enabling GGUF compatibility and utilizing MLX for Apple Silicon acceleration. Feedback is requested on performance and memory utilization.

This release shifts the core architecture to directly support llama.cpp and the GGUF format, introduces MLX acceleration for Apple Silicon, and fixes case handling for the nomic-embed-text model.

This release transitions Ollama's architecture to directly support llama.cpp, enabling GGUF compatibility and MLX acceleration on Apple Silicon. A known issue is that `nomic-embed-text` now enforces lowercase input.

This pre-release version overhauls the architecture to directly support llama.cpp, enabling GGUF compatibility and leveraging MLX for Apple Silicon acceleration. Users are encouraged to provide feedback on performance and stability.

This pre-release updates the architecture to directly support llama.cpp and the GGUF format, while introducing MLX acceleration for Apple Silicon inference. Feedback is requested on performance and stability.

This pre-release updates the core architecture to directly support llama.cpp, enabling GGUF compatibility and leveraging MLX for Apple Silicon acceleration. Users are encouraged to provide feedback on performance and stability.

This pre-release version overhauls the architecture to use llama.cpp directly, enabling GGUF support and leveraging MLX for Apple Silicon acceleration. Users are encouraged to provide feedback on performance and stability.

Ollama 0.30 introduces significant performance and compatibility improvements by integrating llama.cpp, broadening hardware support, and adding new model capabilities.

This pre-release updates the architecture to directly support llama.cpp and the GGUF format, while introducing MLX acceleration for Apple Silicon inference.

This pre-release updates the architecture to directly support llama.cpp, enabling GGUF compatibility and leveraging MLX for Apple Silicon acceleration. Users are encouraged to test performance and stability.

This release focuses on stability and improvements within the MLX backend, including refined model pushing and fixes for inference timeouts and metallib leakage.

This release focuses on hardening update flows and refining model push behavior within the MLX integration.

This release focuses on improvements to the Ollama server caching and the desktop launch experience, including plan-aware model gating and disabling Claude Desktop launch.

This release introduces significant performance improvements via API response caching and refines the launch integration management workflow. The default behavior of `ollama launch` has been updated regarding Claude Desktop integration.

This release introduces Gemma 4 MTP speculative decoding support for Macs, significantly boosting performance for the Gemma 4 31B model on coding tasks, alongside underlying threading fixes and a Go version bump.

This release introduces Gemma 4 MTP speculative decoding support for Macs, significantly boosting performance for the Gemma 4 31B model, alongside general threading fixes and a Go version bump.

This release introduces support for Claude Desktop integration with Ollama and enhances stability by fixing Windows gateway timeouts and hardening Metal initialization.

This release introduces new features like sourcing featured models from an experimental endpoint and adds launch support for the Claude application. It also includes a fix for OpenClaw gateway timeouts on Windows.

This release introduces model batching support and fixes several issues related to tokenization and desktop application startup behavior. It also includes support for NVIDIA TensorRT Model Optimizer import.

This release introduces model batching support and TensorRT Model Optimizer import for the mlx backend. It also includes several bug fixes related to tokenization and desktop application startup behavior.

This release introduces model batching support and adds NVIDIA TensorRT Model Optimizer import capability. Several minor bugs related to tokenization and desktop app session handling were also resolved.

This release introduces support for NVIDIA TensorRT Model Optimizer import within mlx and fixes an issue related to multi-regex BPE offset handling in the tokenizer. It also includes performance improvements by batching the sampler across multiple sequences in mlxrunner.

This release introduces two new models: NVIDIA's Nemotron 3 Omni and Poolside's Laguna XS.2.

This release introduces flexibility in the API by allowing "max" for the think parameter and improves OpenAI response mapping for reasoning effort.

This release introduces structured outputs and ollama cloud support, alongside updating the web search mechanism to use bundled OpenClaw.

This release introduces structured outputs and ollama cloud support, alongside updating the web search mechanism to use bundled OpenClaw.

This release introduces kimi CLI integration and includes several performance and correctness fixes for MLX models and server formatting logic.

This release introduces kimi CLI integration and includes several performance and correctness fixes across MLX models and server formatting logic.

This release introduces Copilot CLI integration and support for the hermes model within the launch command, alongside several fixes to configuration handling during launch.

This release introduces Copilot CLI integration and support for the hermes model within the launch command, alongside several fixes related to launch configuration handling.

This release introduces Gemma4 support on the MLX backend and updates the ROCm version to 7.2.1 on Linux. It also includes various fixes and improvements for MLX operations and Gemma4 rendering.

This release primarily updates the ROCm dependency to version 7.2.1 on Linux and fixes a quality regression in specific Gemma model configurations.

This release focuses on improving Gemma 4 and parallel tool calling capabilities, alongside general application bug fixes and documentation updates for the Hermes Agent.

This release includes documentation updates, fixes for parallel tool call indexing, and UI adjustments for image attachment validation upon model change. The Gemma4 renderer was also updated.

This release introduces documentation for Hermes agent integration and includes several bug fixes related to model parsing, Gemma4 handling, and UI validation upon model switching.

This release introduces setup for openclaw channels and improves command-line interaction, alongside refining safetensors handling and error reporting.

This release focuses on improving the command line interface, refining launch configurations for specific models (glm-5.1, gemma4), and enhancing setup for openclaw and opencode integration. It also includes several minor bug fixes across the application.

This release introduces OpenClaw channel setup via `ollama launch openclaw` and enables flash attention and tool call repair for Gemma 4 models. A bug fix was also implemented for the `/save` command.

This release introduces OpenClaw channel setup for integrating messaging platforms like WhatsApp and Telegram via `ollama launch openclaw`, enables flash attention for Gemma 4, and fixes a bug in the /save command.

This release focuses on performance improvements for MLX (M5 with NAX) and Gemma4 (flash attention), alongside minor fixes for model creation.

This release focuses on performance improvements for M5 models via NAX integration and enables flash attention support for gemma4.

This release focuses on performance improvements for MLX (M5 with NAX) and Gemma4 (flash attention), alongside fixes for model creation paths and safetensor loading.

This release includes improvements to Gemma 4 Tool Calling, adds the latest models to the Ollama App, and fixes issues with launching the OpenClaw TUI.

This release updates the default application home view to use the new chat interface by default. It includes minor changes related to the application's user interface.

This patch release introduces new benchmarking capabilities and resolves several parsing and build issues related to gemma4 and ROCm builds.

This release introduces performance improvements via flash attention for gemma4 and fixes several parsing and build issues related to argument handling and ROCm compilation.

This release introduces the new Gemma 4 model family variants (E2B, E4B, 26B, 31B) and enhances tokenizer capabilities with SentencePiece-style BPE support.

This release introduces the new Gemma 4 model family variants (E2B, E4B, 26B, 31B) and enhances tokenizer capabilities with SentencePiece-style BPE support.

This release introduces improvements to KV cache handling, adds a web search plugin to `ollama launch pi`, and resolves several model loading and parsing bugs across different architectures.

This release introduces a warning for small context lengths and improves launch logic for VS Code integration, alongside various CI and TUI updates.

This release introduces changes to the launch command behavior, improves VS Code path detection, and includes several CI/build hardening updates.

This release introduces changes to the launch command behavior, updates the TUI title handling, and improves CI build processes for MLX and CUDA.

This release focuses on stability improvements, including fixing a memory leak in mlx and adjusting settings for the Grok model on ggml. It also updates VS Code documentation and hides the VS Code launch option.

This release introduces debug request logging, improves KV cache sharing in mlxrunner, and fixes several stability issues including desktop app loading hangs and mlxrunner deadlocks.

This release introduces debug request logging and improves MLX performance with better cache sharing and new format imports. Several stability fixes were also implemented across the desktop app, MLX runner, and CI.

This release introduces significant performance and feature enhancements for MLX backend, including model eviction, quantized embeddings, and fast SwiGLU. It also includes a fix for the web_search legacy path in the cloud proxy.

This release introduces checks for npm and git installation prerequisites for OpenClaw and significantly speeds up local Claude Code execution. Several minor bugs related to model launching and package registration were also fixed.

This release introduces significant performance and feature enhancements for the MLX backend, including model eviction, quantized embeddings, and fast SwiGLU. It also includes a fix for the web_search legacy path in the cloud proxy.

This release focuses on improving the benchmarking tool and adding stability fixes for launch commands, particularly concerning headless mode and systemd availability.

This release focuses on improving the benchmarking tool and adding stability fixes for launch commands, particularly concerning headless mode and systemd availability.

This release introduces documentation for `reasoning_effort` support in the OpenAI-compatible API and includes several fixes related to cloud model handling and launch command integration.

This release introduces documentation for reasoning_effort support in the OpenAI-compatible API and fixes several issues related to cloud model handling and launch command integration.

This release focuses on stability and performance improvements, including fixes for GLM tool calls, localhost handling, and updates to MLX and ROCm support. It also addresses an issue where resetting defaults disabled auto-updates.

This release focuses on stability and fixes, including repairs to GLM tool parsing, localhost handling, and cloud proxy stream disconnects, alongside build improvements for Windows and MLX.