The latest AI news and updates of Ollama — Open-source tool for running, managing, and serving large language models locally. Covering Ollama's latest product updates and company news from the past 90 days.
OllamaOct 21Ollama added Recursive’s Flint, a 40B-parameter mixture-of-experts model designed for local agentic coding. The model features 2.5 billion active parameters per token and a 262K-token context window. It achieves over 125 tokens per second on an M4 Pro MacBook Pro and 150 tokens per second on an NVIDIA RTX Spark, accessible via ollama run flint.
OllamaOct 19Ollama adds MiniMax M3.1 Flash to Ollama Cloud, featuring a 1M-token context window and image understanding for coding and agentic workflows. The model is hosted in the US with zero data retention. Access it via the command line with ollama run minimax-m3.1-flash:cloud or integrate it into Claude Code using ollama launch claude --model.
OllamaOct 16Ollama added NVIDIA’s Nemotron 3.5 Bolt 74B-A9B, a vision-language model designed for local agentic tasks. The model uses native 2-bit quantization to fit within 24GB of memory while supporting a 256K-token context window. It features 9 billion active parameters per token and is available for local execution via the ollama run nemotron-3.5-bolt command.
OllamaOct 16Ollama now hosts Meta's Muse Spark Open, a coding and tool-use model featuring a 1M-token context window. The model runs via the cloud command ollama run muse-spark-open:cloud and integrates with Claude Code using ollama launch claude muse-spark-open:cloud. Access is available to Max, Teams, and Pro users.
OllamaOct 15Ollama added Liquid AI's LFM2.5-24B-A2B to its library, enabling local execution of the 24B-parameter reasoning model. The model features 2B active parameters per token, a 131K-token context window, and competitive performance on agentic tasks compared to larger dense models. It is available for immediate use via the ollama run lfm2.5:24b command.
OllamaOct 3Ollama added Cloudflare's Clef (27B) and Clef Flash (9B) decision models to its library. These models perform typed judgments like image classification, bug report labeling, and ticket routing by returning structured choices and probabilities. The models are accessible via the ollama pull command and operate through the /v1/systemone API endpoint for local, non-autoregressive decision-making.
OllamaSep 30Ollama added support for Jev-style decision models, enabling fast, typed judgments like ticket triage and content moderation entirely locally. The update introduces a /v1/systemone API endpoint in Ollama 0.35, supporting Bespoke Labs' Nimble 9B and Together AI's Tev1 models. These models return structured choices, scores, and probabilities for software to consume directly without network latency.
OllamaSep 24Ollama now allows the purchase of usage credits for paid cloud models without requiring a monthly subscription. This pay-as-you-go model draws from a prepaid balance to cover token usage at published rates. Credits unlock access to all cloud-hosted models, with no service fees or recurring commitments required for usage.
OllamaSep 13Ollama cloud models are now accessible through Amp's BYOK model routing. This integration allows routing to Ollama's hosted models with no limits or fees for BYOK, supporting the creation of remote agents controllable from anywhere.
OllamaSep 11Ollama has begun rolling out DeepSeek-V4.1-Flash on its cloud platform, starting with Max and Team account holders. The model features a 1-million-token context window and native visual understanding. Ollama is currently adding capacity to extend access to all subscribers.
OllamaSep 11Ollama 0.34 enables ChatGPT Desktop to run on local and cloud-hosted Ollama models. The integration supports plugins, web search, and computer use out of the box. Model selection is customizable within the app’s settings, and the setup process runs via the command line using the ollama launch codex-app command.
OllamaSep 6Ollama introduced off-peak token rates for DeepSeek-V4-Flash and Pro models on its cloud service. These models are now half price outside the 12:00 to 18:00 UTC weekday peak window and all day on weekends. The service, hosted in the US and Europe with zero data retention, plans to expand off-peak pricing to additional models soon.
OllamaSep 1Ollama transitioned its Pro, Max, and Team cloud plans to transparent per-token pricing, replacing previous GPU-time billing. Each plan now includes a monthly pool of usage credits, such as $60 for the $20 Pro tier and $1,000 for the $500 Team plan. The Team plan is now generally available for unlimited users with shared credits and no service fees.
OllamaAug 29Ollama has fully rolled out Z.ai's GLM 5.3 and GLM 5.3 Flash models on its cloud service. Hosted in the US and Europe with zero data retention, these models support coding harnesses like Claude Code and OpenCode. Both models are accessible via API key, with adjustable reasoning effort settings available for the flagship GLM 5.3.
OllamaAug 26Ollama v0.33 adds Claude Desktop support, connecting the application as a third-party gateway provider. This integration brings local models and Ollama cloud models into the Claude Desktop interface. The update includes support for auto mode and web search, while maintaining Ollama’s Zero Data Retention policy for all requests.
OllamaAug 25Ollama now hosts IBM Granite 4.2, a family of open foundation models in 3B, 8B, and 30B parameter sizes. These models feature a 128K context window, multilingual support, and a built-in thinking mode for reasoning. Licensed under Apache 2.0, the models are trained with governance, risk, and compliance evaluations specifically for enterprise agent applications.
OllamaAug 15Ollama has fully rolled out DeepSeek-V4-Pro (0813) on its cloud platform for Pro and Max subscribers. This 1.6T-parameter Mixture-of-Experts model features a 1-million-token context window and three distinct reasoning modes for logical analysis. The service is hosted in the US with a Zero Data Retention policy, accessible via the ollama run deepseek-v4-pro:cloud command.
OllamaAug 14Ollama will support Z.ai's upcoming GLM-5.3 model upon its open release. The model, built on a 743B base, features specialized capabilities for coding, agentic workflows, and cybersecurity. This integration brings the model's post-trained performance to local environments, supporting autonomous code generation and defense tasks.
OllamaAug 12Ollama now functions as a provider for GitHub Copilot in JetBrains IDEs. This integration enables provider configuration and model selection directly within the JetBrains experience. Local models managed by Ollama now power Copilot chat and code suggestions, offering an alternative to cloud-based model options.
OllamaAug 11Ollama partners with NVIDIA to launch Nemotron 3.5 Lightning, a 30B-parameter mixture-of-experts model with 3.6B active parameters. The model targets high-volume agentic workflows, providing efficient reasoning for long-running tasks. This integration makes the open model available for local execution, supporting autonomous agent development.
OllamaAug 10Ollama adds support for Meta’s Muse Glimmer 30B model, enabling local agent workflows on Apple Silicon. The integration uses Ollama’s MLX engine with DFlash for 1.5x–1.8x faster performance and adds native image input support for coding agents. The model integrates with tools like Claude Code and Codex, featuring controllable reasoning strength settings.
OllamaAug 8Ollama rolled out DeepSeek-V4-Flash-0731 as the new default model for its cloud inference service. The update delivers over 120 output tokens per second with zero data retention in the US and Europe. Pro and Max plan subscribers receive generous usage for long-running sessions with coding agents like Claude Code, OpenCode, and Hermes Agent.
OllamaAug 5Ollama is scaling its cloud capacity in the US and Europe following record-breaking token usage for the DeepSeek-V4-Flash-0731 model. The model now runs on Ollama’s cloud with performance exceeding 100 tokens per second. Ollama maintains zero data retention for all requests, ensuring user data remains private during inference.
Frequently asked questions