Ollama
Local LLM runner releases. Ships new model support and inference improvements.
Release history
Ollama 0.40.9
Llama.cpp dependency updated to improve model inference performance and compatibility
Ollama 0.40.7
Ollama 0.40.7 adds Perplexity's 27B decision model, fixes OpenAI-compatible streaming with reasoning models, and cancels remote requests on client disconnect
Ollama 0.40.6
Ollama 0.40.6 adds Q2_0 quantization support, fixes Windows runner cleanup, and improves cloud integration with Clef
Ollama 0.40.5
Ports Metal custom kernels to CUDA, relevant for Nvidia GPU users running mlx-style workloads
Ollama 0.40.4
Updates bundled llama.cpp for decision-model and EmbeddingGemma 2 support, plus faster manifest reuse in inference routes
Ollama 0.40.3
Cleans up warning messages for Claude and Codex launch commands; No functional change for most setups
Ollama 0.40.2
Ollama 0.40.2 upgrades downloaded models for llama.cpp compatibility while keeping backups, and fixes duplicate entries in model listings
Ollama 0.40.1
Adds proxy cloud usage/balance APIs and fixes a Windows memory-read bug past 2GiB in llama backend
Ollama 0.35.1
Now allows ten web searches per response. Includes llama.cpp and MLX version bumps
Ollama 0.40.0
MLX-compatible models run on Apple Silicon
Ollama 0.34.4
Fixes intermittent model-not-found errors on the server side
Ollama 0.34.3
GET /api/show now advertises each model's thinking controls and default values
Ollama 0.34.2
Llama.cpp updates only
Ollama 0.34.1
Token repeat limit raised to 100. MLX runner checks free memory before loading models and evicts prefix cache from active conversations
Ollama 0.34.0
Ollama adds ChatGPT Desktop integration, improves structured output on Apple Silicon, and supports OpenAI-compatible client tool search
Ollama 0.33.3
Honors GGUF model-defined default parameters; Updates MLX and llama.cpp
Ollama 0.33.2
Ollama fixes dark mode support, prevents duplicate macOS instances, and stops interrupting Claude Desktop requests during model catalog updates
Ollama 0.33.1
Ollama 0.33.1 adds Qwen3.8 Flash support, structured output for MLX models, and fixes Metal GPU timeouts during model loading
Ollama 0.33.0
Claude Desktop integration: toggle individual Ollama models on and off from the menu bar, choose models directly within Claude
Ollama 0.32.15
Model metadata caching halves time-to-first-token, parser error handling fixed, Qwen system messages normalized
Ollama 0.32.11
Ollama added DeepSeek Harness and Meta Muse Code agent support, plus web search in OpenAI-compatible Responses API
Ollama 0.32.10
Ollama 0.32.10 changes default repeat penalty to 1.0 for faster speculative decoding, speeds NVFP4 prefill%, and fixes blob verification on shared digests
Ollama 0.32.9
Ollama adds NVIDIA's Nemotron 3.5 Lightning, a 30B mixture-of-experts model optimized for always-on agent deployment
Ollama 0.32.8
Muse Glimmer model now available across all platforms, optimized for coding agents and personal assistants with MLX performance on Apple Silicon
Ollama 0.32.7
Muse Glimmer, Meta's 30B multimodal agent model, now runs locally on Ollama with Apple Silicon optimization and image input support
Ollama 0.32.6
Ollama 0.32.6 speeds up Qwen3.5 on Apple GPUs via speculative decoding, fixes OpenAI API streaming format compliance, and improves TUI usability
Ollama 0.32.4
Ollama 0.32.4 adds Laguna support for Apple GPUs, fixes Qwen3 MoE expert quantization, and improves speculative decoding performance
Ollama 0.32.3
Ollama 0.32.3 fixes stalled downloads, expands GPU support to Windows ARM64 and B200, and adds chat/tool-calling for Laguna 2.1 models
Ollama 0.32.2
Ollama 0.32.2 adds agent skills system, fixes Anthropic thinking-block handling, updates llama.cpp and MLX backends, and enables unlimited tool rounds for cloud models
Ollama 0.32.1
Ollama 0.32.1 fixes MLX model cache leaks, improves Gemma 4 tool calling, and adds working directory context to agents
Ollama 0.32.0
Adds Qwen3.5 parser and renderer selection; Agent UI warns before loading old agent models
Ollama 0.31.2
Flash attention now works on older NVIDIA GPUs, vision model offloading improved, and non-UTF-8 model paths fixed
Ollama 0.31.1
Gemma 4 nearly 90 percent faster on Apple Silicon via multi-token prediction
Ollama 0.30.12
Fixed tool call parsing to handle JSON strings with braces correctly, and updated llama.cpp and MLX dependencies
Ollama 0.30.11
Ollama 0.30.11 fixes GPU detection on hybrid Windows systems, improves speculative decoding, and corrects memory reporting for partially offloaded models
Ollama 0.30.10
Command and Northfamily models now run on Apple Silicon via MLX engine, llama.cpp updated to build 9672
Ollama 0.30.9
Ollama 0.30.9 adds Cohere2Moe support, fixes token output limits in coding agents, and validates message sizes against context windows
Ollama 0.30.8
Ollama 0.30.8 improves prompt caching efficiency, fixes provider selection bugs, and hardens MLX inference stability with snapshot creation
Ollama 0.30.7
Ollama adds Hermes Desktop, a native UI for managing Hermes agent conversations and integrations via ollama launch
Ollama 0.30.6
Ollama adds Gemma 4 QAT models for lower memory use, integrates Oh My Pi coding agent, improves Apple Silicon quantization
Ollama 0.30.3
Ollama added Gemma 4 12B model support, expanding locally-runnable open-weight options for builders
Ollama 0.30.2
Auto-install Cline CLI; Qwen code integration added; Template logging improved for troubleshooting
Ollama 0.23.4
Ollama launch now handles vision models with image inputs. Local image paths work in Claude tool results
Ollama 0.30.0
Ollama now builds directly on llama.cpp with GGUF format support and MLX acceleration for Apple Silicon inference