Skip to main content
Local LLM Guide

News

Ollama 0.32: Local Model Runner or AI Agent?

By 6 min read

For most of its life, Ollama had one job on this site: it was the quiet backend. You pulled a model, you pointed something at the API, and the interesting decisions happened somewhere else. Version 0.32, released on 11 July 2026, changed the front door.

In version 0.32, typing ollama with no arguments opens an interactive agent instead of help text. It offers chat, coding, web search, and delegated work. The release's banner names a hosted model. These are the 0.32 launch defaults; check the selected model and server settings in your current installation.

That last detail is the whole story, so let us take it slowly.

What actually changed in 0.32.0

Four things, per the release notes:

  1. The bare ollama command launches an interactive agent. The startup banner advertises chat, code, web search, and delegated work.
  2. The Codex App integration was renamed to ChatGPT. Use ollama launch chatgpt, with --restore to return to your usual ChatGPT profile.
  3. The ollama launch menu was simplified to the most popular integrations. The others remain reachable through ollama launch directly.
  4. Older agent models now warn before launching. CodeLlama, Qwen2.5 and Qwen2.5-coder, the Llama 3.x line, Mistral, StarCoder, and the base DeepSeek-R1 tags all raise a deprecation warning first.

What is conspicuously absent from that list matters as much as what is on it. The release notes record no change to ollama run, and no change to any other model-running subcommand. If your workflow is a script, a systemd unit, a Docker container, or anything else that calls a subcommand explicitly, nothing in 0.32.0 was aimed at you. The new behaviour attaches to the bare invocation only.

The part worth pausing on

The agent banner in the official 0.32.0 notes names glm-5.2:cloud as the model behind the new experience. The :cloud suffix is not decoration. It denotes a model served from Ollama's hosted infrastructure rather than from your GPU.

Version 0.32.1, five days later, added a further tell: agent web search and web fetch now instruct the user to run ollama signin when authentication is required. A local model runner does not need you to sign in. An account-backed cloud service does.

The release makes the execution path especially important. Installing Ollama alone does not establish that a request is local. ollama run with a locally downloaded model is a local inference path; a hosted model and web tool have different destinations. Check both before sending sensitive material.

If you want the old assumption to hold, the practical guidance is unchanged in shape but now needs stating explicitly:

  • Call subcommands rather than the bare command. ollama run <model> against a model you pulled locally is a local inference path.
  • Pin models locally rather than accepting a cloud-tagged default.
  • Check the destination of each model and tool request. Sign-in alone does not tell you where inference happens: it can also authorize downloading private models.

The current Ollama FAQ provides a direct local-only setting: set OLLAMA_NO_CLOUD=1 in the Ollama server's environment and restart it. Alternatively, add "disable_ollama_cloud": true to the existing JSON object in ~/.ollama/server.json, preserving any other settings, then restart. This disables Ollama cloud models and web search; it does not restrict unrelated apps or external agent tools.

Ollama 0.35 also adds a separate local decision endpoint. A classifier uses /v1/systemone with a compatible GGUF model, rather than a cloud agent or ordinary chat endpoint. Follow the Nimble ticket-sorting recipe to make a local request and inspect its result.

The 0.32 point releases, briefly

The series moved fast, and one release should be skipped outright.

  • 0.32.1 (16 July). Improved Gemma 4 tool calling and multi-turn reasoning, including more reliable continuations after a tool response. Fixed a recurring MLX model cache leak that could grow memory use across requests. MLX text model loading now respects OLLAMA_LOAD_TIMEOUT. The agent now receives the current working directory for project context.
  • 0.32.2 (20 July). Withdrawn. Use 0.32.3 or newer.
  • 0.32.3 (23 July). Fixed model downloads that stalled before sending data. Restored Claude Code channels and fixed Anthropic thinking streams. Added CUDA on Windows ARM64 and B200 support through CUDA 12, and lowered memory use on Linux CUDA and ROCm integrated GPUs. Added chat, thinking, and tool calling for Laguna 2.1.
  • 0.32.4 (25 July). Laguna on Apple GPUs through MLX. Fixed Qwen3 MoE decoding for differently quantized experts, with a faster packed gate and up projection that Ollama reports at roughly 4 to 9 percent on an M5 Max.
  • 0.32.5 (27 July). Fixed an MLX Metal bug that could reduce output quality for NVFP4 models.
  • 0.32.6 (4 August). Qwen3.5 is faster on Apple GPUs, because the MLX engine now uses the model's MTP head for speculative decoding automatically. OpenAI-compatible streaming was brought in line with the upstream wire format. ollama run kimi-k3 now offers the cloud tag rather than failing for a cloud-only model that publishes no default tag.

One historical regression deserves a callout rather than a bullet. 0.32.6 temporarily removed experimental image generation. Its release note advised staying on 0.32.5 for that feature at the time. That is guidance about those two releases, not a recommendation to install an older version today; check the current release's image support before changing your setup.

About that Apple Silicon speed claim

You will see a figure circulating that Ollama generates Gemma 4 tokens nearly 90 percent faster on Apple Silicon, averaged across a coding-agent benchmark, by leaning on multi-token prediction. That number is Ollama's published result.

The figure belongs to version 0.31.1, released 30 June 2026, and to Ollama's selected coding-agent benchmark. It is not a universal speed increase for every Gemma file, device, or conversation. Compare the same target model, task, and cache state on your own machine.

The mechanism is at least easy to describe: Ollama auto-tunes how many tokens to draft as it runs, so the acceleration is on by default, needs no configuration, and Ollama states it does not change model output.

So which is it?

Both, and that is the honest answer rather than a hedge.

Ollama in August 2026 is still the same local inference engine underneath. The MLX and llama.cpp work across the 0.31 and 0.32 series is exactly the sort of unglamorous backend improvement that made it worth recommending in the first place. The release notes describe specific fixes and engine changes; their effect depends on the model and hardware you use.

What changed is that the project now also ships an agent product, and the agent product is the thing you meet first. For a general developer audience that is a reasonable default. For an audience that came to Ollama specifically because it was the boring local option, the default now points somewhere else, and the burden is on the user to steer away from it.

That is worth knowing before you upgrade. It is not a reason to avoid upgrading.

Sources