Ollama
Download and run open AI models like Llama, Gemma and Qwen on your own computer
- Made by
- Ollama
- Category
- Model
- Last change
- Nov 3, 2026
Changelog
November 2026
- New feature
Ollama now offers full zero data retention for its cloud-hosted open models, so users can run them on private documents and government data.
What this means for you: Ollama's cloud-hosted open models now come with full zero data retention, meaning your inputs are not kept. That makes them usable on private documents and government data.
Why it matters: Ollama says open models were the first to do sensitive work like private documents and government data, and it wants them to be the default for enterprises.
My organization handles private documents, and I just learned Ollama now offers full zero data retention for its cloud-hosted open models. Help me write a list of questions to ask my IT team before we use it, covering which documents are involved and who should approve. Our documents are: [describe].
- New feature
Ollama's launch command for Claude Code now applies each model's output token limit, so cloud models no longer hit the 128k default cap.
What this means for you: If you use Claude Code through Ollama's launch command, cloud models can now produce longer answers. The launch command applies each model's own output limit instead of the old 128k default cap.
How to use it: Update Ollama to v0.40.14, then start Claude Code with the ollama launch claude command.
October 2026
- Availability
Ollama now runs Liquid AI's LFM3 models, including the lfm3-audio and lfm3-omni text-to-speech models, on Linux and Windows as well as Apple silicon.
What this means for you: The LFM3 models from Liquid AI, including the lfm3-audio and lfm3-omni models that turn text into speech, now run in Ollama on Linux and Windows computers. Before this, they ran only on Macs with Apple silicon.
Why it matters: It follows the Oct 27 entry, when Ollama first added LFM3, a family of small text, vision, audio, omni and decision models.
How to use it: Update Ollama to v0.40.13 on Linux or Windows. Then run the LFM3 models locally, including lfm3-audio and lfm3-omni for text-to-speech.
I just learned Ollama can now run Liquid AI's LFM3 models, including lfm3-audio and lfm3-omni for text-to-speech, on Windows and Linux. Help me set this up: how to update Ollama, which model to try first, and a simple test. My computer is [Windows or Linux] with [amount] of memory.
- New model
Ollama now offers LFM3 from Liquid AI, a family of small text, vision, audio, omni and decision models with a 256K context window, to run locally.
What this means for you: Ollama, an app for running AI models on your own computer, now offers Liquid AI's LFM3 family. It includes small models for text, vision, audio, omni (several input types in one model) and decision-making. The text model fits in about 1 GB of memory at 4-bit, and all of them have a 256K context window, meaning how much text they can handle at once.
How to use it: The LFM3 models are available in Ollama to run locally. The text model needs about 1 GB at 4-bit, and the family can be fine-tuned on a single GPU.
Ollama now offers Liquid AI's LFM3 models for text, vision, audio, omni and decision tasks, and the text model fits in about 1 GB. Explain in plain words which LFM3 model I should try first and what to test. My computer has [amount] of memory and I want to [goal].
- New feature
Ollama can now generate speech on macOS, playing text or replies aloud, and adds the d3 decision models in 2B to 27B sizes.
What this means for you: Ollama runs AI models on your own computer. On a Mac it can now read text or a chat reply out loud as the audio is generated. It also adds d3, models that take a piece of text, an image or structured data and answer yes or no, classify it, or score options with probabilities. For example, given "A customer says their card was charged twice," it can answer whether it is a billing issue.
Why it matters: It comes six days after Ollama added Perplexity's pplx-decider, a 27B decision model, on Oct 18.
How to use it: Speech works on macOS. In Terminal, run ollama run phonon "Hello from Ollama." to hear it. For d3, run ollama run d3 with your text and an --ask question. Sizes are 2B, 4B, 9B and 27B. This is Ollama v0.40.11.
I just learned that Ollama's new d3 models can read a piece of text and answer a yes or no question or classify it, like deciding whether a customer message is a billing issue. Help me design a simple sorting setup for [type of messages I receive], with categories, example questions to ask, and how to check the answers.
- New model
Ollama adds NVIDIA's Nemotron 3.5 Bolt E4B, a small vision-language model with a 128K-token context window and a speculative decoding draft.
What this means for you: Ollama lets you download and run AI models on your own computer. The new Nemotron 3.5 Bolt E4B is a small model from NVIDIA that can talk with you, look at images and use tools. A 128K-token context window (the amount of text it can keep in mind at once) lets it work with long conversations or documents.
How to use it: After installing Ollama v0.40.10 or later, run this command in your terminal: ollama run nemotron-3.5-bolt:e4b
- New model
Ollama now offers NVIDIA's Nemotron 3.5 Bolt E4B 4GB, a small model for local assistants, interactive characters and robots that use tools.
What this means for you: Ollama now offers a small open AI model from NVIDIA that runs on your own computer. It is built for local assistants, interactive characters and robots that carry on a conversation and use tools to get things done.
How to use it: The model is called Nemotron 3.5 Bolt E4B 4GB. Find it in Ollama, or run: ollama run nemotron-3.5-bolt:e4b
- Availability
Ollama now runs the Flint coding model on Linux and Windows as well as macOS.
What this means for you: Flint, a coding model you run on your own computer through Ollama, now works on Linux and Windows. Before this it ran only on Mac. A coding model is an AI that writes and fixes software code.
Why it matters: Flint first appeared in Ollama on Oct 21, a day earlier, and only on macOS. This release extends it to the other two systems.
How to use it: Update Ollama to version v0.40.9, then run Flint on Linux or Windows the same way you would on a Mac.
- New model
Ollama now offers Flint from Recursive SI, a roughly 40B parameter open coding model built to run on-device, usable with Claude Code.
What this means for you: Flint is a coding model that runs on your own computer through Ollama, so your code stays on your machine. It is an open model (the weights, the model's learned settings, are public) of roughly 40B parameters. It can also power Claude Code, Anthropic's coding tool, in place of an online model.
How to use it: Run `ollama run flint` to chat with it. To use it with Claude Code, run `ollama launch claude --model flint`. Ollama reports more than 125 tokens per second (tokens are small chunks of text) on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark, on short contexts.
I want to try Flint, a roughly 40B open coding model that Ollama can run on my own computer, including with Claude Code. Walk me through whether my machine can handle it and how to set it up step by step. My computer: [model, RAM, operating system].
- New model
Ollama now offers Ornith 1.5 122B, which users can run locally on their own device.
What this means for you: Ornith 1.5 122B is an AI model you can download and run on your own computer through Ollama, so your chats stay on your device. It is built for coding and for agent tasks, where the AI uses tools and works through several steps on its own. The weights are open under the MIT license, meaning the model files are free to download and use.
I want to run an AI model locally with Ollama and just learned Ornith 1.5 122B is available, with open weights under the MIT license and a focus on coding and agent tasks. Help me work out whether my computer can run it and what to try first. My computer: [model, memory, graphics card].
- New model
Ollama adds pplx-decider, a 27B decision model from Perplexity fine-tuned from Qwen3.8-27B.
What this means for you: Ollama, the app that runs AI models on your own computer, now offers pplx-decider. It is a 27B decision model from Perplexity, built by fine-tuning (further training) Qwen3.8-27B. The 27B means it has 27 billion parameters, which is a measure of its size.
How to use it: Update Ollama to version v0.40.7.
- New model
Ollama Cloud now offers DeepSeek-V4.1-Flash-1017, an updated DeepSeek model for agentic tasks, hosted in the US with zero data retention.
What this means for you: Ollama Cloud now runs DeepSeek-V4.1-Flash-1017, an updated DeepSeek model that is strong at agentic tasks, where the AI works through several steps on its own. It is hosted in the US with zero data retention, meaning your data is not kept after the request.
How to use it: In a terminal, run: ollama run deepseek-v4.1-flash:1017-cloud. To use it with Claude Code, run: ollama launch claude --model deepseek-v4.1-flash:1017-cloud. For OpenCode, run: ollama launch opencode --model deepseek-v4.1-flash:1017-cloud.
- New model
Ollama Cloud now offers Reflection AI's Beam, a 1M-token context model for software engineering, free for a limited time and hosted in the US.
What this means for you: You can try Beam, a model made for software engineering and long-running agent tasks, through Ollama's cloud at no cost for a limited time. It reads up to 1 million tokens (small pieces of words) at once, so it can take in very long material.
Why it matters: Beam is Reflection AI's first open-weight model, and Ollama is a day 0 partner, so it is available the day it was released.
How to use it: In a terminal, run: ollama run beam. To use it inside coding tools, run ollama launch claude --model beam:cloud, ollama launch codex --model beam:cloud, or ollama launch opencode --model beam:cloud. It is hosted in the US with full ZDR (zero data retention), and it is free for a limited time only.
Ollama Cloud is offering Reflection AI's Beam free for a limited time. It has a 1M-token context and is built for software engineering. Help me plan a test of it on my own project: what to give it, what to ask, and how to judge the result. My project is: [describe project].
- New model
Ollama now offers Liquid AI's LFM2.5 24B, a local model for laptops and PCs that handles multi-step tasks in agent tools.
What this means for you: Liquid AI's LFM2.5 24B is a model you can run on a laptop or PC. It is built to work through multi-step tasks inside popular agent tools. It is a reasoning model, and its weights are open on Hugging Face.
How to use it: In Ollama, run ollama run lfm2.5:24b. The model page is ollama.com/library/lfm2.5:24b.
I want to try Liquid AI's LFM2.5 24B, a local model that Ollama now offers for multi-step tasks in agent tools. Help me decide whether it fits my setup and list three multi-step tasks I could try first. My computer is: [chip, RAM], and my usual tasks are: [describe tasks].
- New model
Ollama's cloud now offers Meta's Muse Spark Open, a coding and tool-use model with a 1M-token context window, for Max, Teams and later Pro users.
What this means for you: Meta's Muse Spark Open is an open-weight model for coding and tool use that runs on Ollama's cloud, not on your computer. It can take in a 1M-token context window, which means very long documents or large projects at once.
How to use it: Run ollama run muse-spark-open:cloud. To use it with Claude Code, run ollama launch claude muse-spark-open:cloud. It is available now to Max and Teams users and is rolling out to Pro users soon.
- New model
Ollama now offers NVIDIA Nemotron 3.5 Super VL, a model combining reasoning and visual understanding for agent workflows.
What this means for you: NVIDIA's Nemotron 3.5 Super VL on Ollama combines reasoning with understanding of images, for use in agent workflows (setups where an AI works through tasks in steps).
How to use it: In Ollama, run ollama run nemotron-3.5-super.
Ollama now offers NVIDIA's Nemotron 3.5 Super VL, which combines reasoning and visual understanding for agent workflows. Help me plan a few tasks where an AI would need to look at images and then reason about them. My work involves: [describe your work and the kinds of images you handle].
- New model
Ollama now offers NVIDIA Nemotron 3.5 Bolt 74B-A9B, a vision-language model that fits in 24GB of memory for local agentic tasks.
What this means for you: You can run a model on your own computer that understands both pictures and text, and it is built to handle everyday multi-step tasks (an "agent" is an AI that works through steps for you). It is compressed to fit in 24GB of memory, so it can run while your other apps stay open.
Why it matters: Ollama added the Nemotron 3.5 Super and the smaller Nemotron 3.5 Bolt on Oct 14. This version is the Bolt model squeezed down to fit in 24GB.
How to use it: Open a terminal and type: ollama run nemotron-3.5-bolt
I use Ollama to run AI models on my own computer, which has [amount] GB of memory. I just learned Ollama now offers NVIDIA Nemotron 3.5 Bolt, a vision-language model sized for 24GB. Help me decide whether my computer can run it, and suggest three everyday tasks to try. My work involves: [describe your work].
- New model
Ollama adds NVIDIA's Nemotron 3.5 Super and the smaller Nemotron 3.5 Bolt, vision-language models for coding, reasoning and tool use.
What this means for you: Two new models you can run on your own computer through Ollama. Both read images and text and are meant for coding, reasoning, tool use and understanding documents. The smaller Nemotron 3.5 Bolt is tuned to fit a 24 GB NVIDIA graphics card for everyday agent tasks, meaning tasks where the model takes steps on its own.
How to use it: Update Ollama to v0.40.5, then run `ollama run nemotron-3.5-super` in the terminal. Nemotron 3.5 Super supports up to 1M tokens of context and an on/off thinking mode.
I'm thinking of running NVIDIA's Nemotron 3.5 Bolt on my own computer with Ollama. It reads images and text and is tuned to fit a 24 GB NVIDIA GPU. Help me check whether my machine can run it and which everyday tasks to try first. My computer: [graphics card, memory, operating system].
- New feature
Ollama now lets users run decision models from the terminal to answer yes/no questions, classify text or score against a rubric.
- New feature
Ollama now upgrades previously downloaded models in the background on first run for better performance, keeping backups of the originals on disk.
- New model
Ollama now offers Google DeepMind's EmbeddingGemma 2, a multimodal embedding model made for consumer devices.
- New model
Ollama now offers Cloudflare's Clef (27B) and Clef Flash (9B) decision models for classifying images, labeling bug reports and routing support tickets.
September 2026
- New feature
Ollama now supports decision models locally through a new /v1/systemone API, for tasks like ticket triage, model routing and content moderation.
- New model
Ollama adds Cloudflare's Clef and Clef Flash multimodal decision models, which accept images alongside text.
- New feature
Ollama adds decision models that return choices, probabilities and scores instead of text, starting with Nimble and Tev1.
- New feature
Ollama on Apple Silicon now runs supported model architectures on the MLX runtime by default, and adds decision models and an embedding model on MLX.
- Pricing
Ollama now lets users add usage credits and pay as they go for paid cloud models without an Ollama subscription.
- New feature
Ollama now shows each model's thinking controls and default setting in the CLI, API and on ollama.com for cloud models.
- New feature
Ollama now offers a first-run setup when running the command, with options to sign in or continue locally, shared with the desktop app.
- New feature
Ollama 0.34 lets the ChatGPT Desktop (Codex) app be configured to use local and cloud Ollama models.
- New model
Ollama Cloud now offers DeepSeek-V4.1-Flash to all plans and pay-as-you-go users, hosted in the US and Europe with zero data retention.
- New model
Ollama is rolling out DeepSeek-V4.1-Flash on its cloud, starting with Max and Team accounts and expanding to all subscribers.
- New feature
Ollama models can now be used directly in ChatGPT Desktop, with setup available from the Ollama app on macOS.
- Pricing
Ollama now charges half price for DeepSeek-V4-Flash and Pro on its cloud outside 12:00 to 18:00 UTC on weekdays and all day on weekends.
- New feature
Ollama's MLX engine now supports images and audio with Gemma 4.
- New feature
Ollama now supports the Muse Code harness, which users can start with the command ollama launch muse using local or cloud models.
- Pricing
Ollama moved its Pro, Max, and Team plans to per-token pricing with a monthly pool of usage credits, while existing plans keep working.
August 2026
- New model
Ollama has fully rolled out GLM 5.3 and GLM 5.3 Flash on its cloud, hosted in the US and Europe with zero data retention.
- New feature
Ollama v0.33 lets users set up Claude Desktop to use Ollama as a third-party gateway provider with one toggle, for cloud and local models.
- New model
Ollama now offers IBM Granite 4.2 in 3B, 8B, and 30B sizes, free to use for research and commercial work.
Related tools
- DeepSeekChinese lab's open-weight models and free chat app
- OpenRouterOne account and one API to use hundreds of AI models from every major lab
- ReplicateRun open AI models in the cloud with one click or one line of code
- HunyuanTencent's open AI models for video, images and 3D
- KimiMoonshot's open-weight models and chat app
- LM StudioA desktop app to find, download and chat with open AI models offline
