Product Marketing Manager

Ollama v0.40.2 upgrades models in the background. 2026-10-09 Ollama released v0.40.2 on 9 October 2026. The release changes how previously downloaded models are handled, upgrading them in the background the first time they are run with the new version. The upgrade is intended to improve performance and compatibility when models run on llama.cpp. Ollama also keeps the original model copy as a backup, so users can downgrade safely if they need to return to the earlier version. For someone making things, the main benefit is a less disruptive upgrade path. Existing models do not need to be manually replaced before testing v0.40.2. Instead, Ollama performs the model upgrade when the model is first run. That reduces the amount of setup work involved when moving an existing local workflow to the new release. The change matters most for creators who already have models downloaded and rely on repeatable local runs. A model that worked with an earlier Ollama version may need a compatible format or runner configuration for llama.cpp. The release notes do not list individual model families or specific performance results, but they identify better llama.cpp performance and compatibility as the purpose of the background upgrade. There is a storage tradeoff. Because Ollama preserves the original copy, upgraded models temporarily occupy additional disk space. This backup allows a safer downgrade, but users with large model collections should check available storage before running many models for the first time after updating. The release notes say that a future Ollama release will remove these backups automatically. For now, users can remove backed-up models manually. The cleanup method requires jq and uses Ollama's local API to find model manifests associated with the llamacpp runner, identify older ggml entries, and remove the matching objects. That command should be used carefully, because it deletes backup data rather than simply hiding it. The release is available on Mina Labs for 0 credits. That makes it possible to use the listed offering without treating the local model upgrade process as a separate purchasing decision. The practical value here is the combination of access and a release that handles model compatibility in the background. Minaxlab would use Ollama v0.40.2 for an existing local workflow where several downloaded models need to be tested against a newer llama.cpp-based runtime. Minaxlab would update Ollama, run each model once, and allow the background conversion to complete before comparing results. Minaxlab would also monitor disk usage during that process and keep the original backups until the updated models had been checked. For a production or shared environment, Minaxlab would test the release on a copy of the workflow first. The backup behavior makes rollback safer, but it does not remove the need to verify outputs, storage requirements, and any scripts that depend on the local Ollama API. After validation, the manual cleanup command could remove no-longer-needed backups until automatic cleanup arrives in a later release.

AI developer digest - Sep 24, 2026. Fabio Pacifici Full-stack developer & educator Local models got more useful this week. Ollama shipped two quality-of-life changes developers have been asking for, Perplexity brought its agentic "Computer" fully on-device to Windows, and there's a paper that quietly reframes how Fabio Pacifici should think about what an LLM is. Let's go. 1. Ollama now streams tool calls. Ollama added streaming responses with tool calling. The old behaviour forced you to wait for the full generation, then parse the JSON to figure out whether the model wanted to call a tool or just return text. The new parser reads each model's template directly to understand tool-call structure as it streams, so your chat applications can render content and fire tool calls in real time instead of stalling at the end of the response. Why it matters: this is the difference between an API that's usable for agent loops and one that feels like a batch job. Streaming tool calls make local agent orchestration feel native - and the parser handles models that weren't even explicitly trained on tool tokens, which is a nice hedge for the wide zoo of open models on Ollama. 2. Turn thinking on and off. Ollama also added a think parameter (and -hidethinking / /set nothink in the CLI) so you can toggle a model's reasoning pass per application. Want the 8B model to mull over multiple viewpoints before answering? Enable thinking. Need sub-200ms latency for a UI autocomplete? Disable it. For scripting, -hidethinking strips the thinking trace and returns only the answer. Why it matters: as reasoning models proliferate, having explicit control over when the "think" pass runs - and hiding it from the output stream when you don't need it - is a real workflow lever. It's a small API detail with big UX impact for anyone building chat or agent interfaces. 3. Perplexity's local agent lands on Windows RTX. Perplexity shipped "Portable Computer" in its Windows app for GeForce RTX and RTX PRO systems (24GB+ VRAM). It's a local version of the Computer agent that plans and executes multistep tasks using on-device models (e.g. Qwen 3.8 27B, post-trained and RTX-optimised), keeps sensitive data on the machine, and only asks before sending anything to the cloud for heavier reasoning. Connectors for Outlook, Drive, Gmail, Slack and GitHub round it out. Why it matters: this is the on-device agentic trend maturing - real multistep agents running fully local, no per-query credits, private by default. For developers it signals where local inference and agent orchestration are converging: the PC as a first-class agent runtime rather than just a cloud client. 4. A probabilistic lens on LLMs (arXiv). "The Probabilistic Structure of Large Language Models" gives a self-contained account of LLMs as probability measures over token sequences - training as maximum-likelihood estimation, generation as sequential simulation of that stochastic process - and ties the asymmetry of KL-divergence to familiar failure modes like hallucination and the gap between statistical plausibility and truth. Why it matters: most working material treats models either as magic or as black boxes. A clean probabilistic framing is genuinely useful for reasoning about why models deviate from the training distribution, and it connects nicely to diffusion models for engineers who think in terms of score functions. Worth the read. What to watch. Local tool-calling and local agents are both getting good enough to build real products on. The interesting question over the next few weeks: which app patterns actually benefit from staying on-device, versus the ones that are better served by cloud reasoning? Building something local? The tooling has never been closer to turnkey. Sources. * Ollama - "Streaming responses with tool calling" - https://ollama.com/blog/streaming-tool - May 28, 2025 * Ollama - "Thinking" - https://ollama.com/blog/thinking - May 30, 2025 * NVIDIA - "Perplexity Portable Computer Is Now Available on Windows, Powered by NVIDIA RTX" - https://blogs.nvidia.com/blog/local-ai-perplexity-windows-pcs/ - Sep 2026 * arXiv - "The Probabilistic Structure of Large Language Models" - https://arxiv.org/abs/2609.25134 - Sep 2026

Ollama v0.34.0 brings its models to ChatGPT Desktop. 2026-09-09 Ollama has released v0.34.0, adding a way to use Ollama models directly in ChatGPT Desktop. The release also updates several parts of its OpenAI-compatible interface. The announcement is documented in the v0.34.0 release notes. Setup is available through the Ollama app on MacOS. The update also improves structured output performance on Apple Silicon, adds support for client tool search and response compaction, and fixes image handling through compacted responses. For people making things, the main change is workflow flexibility. You can keep using ChatGPT Desktop as the place where you write prompts, review responses, and work through a project, while selecting Ollama models for supported tasks. That reduces the need to move text between separate applications or build a custom interface before testing an open model. The release is especially relevant for creators who already use structured output. Structured output is useful when a model needs to return predictable fields instead of free-form prose. A content pipeline might ask for a title, summary, tags, and scene list in a fixed format. A design tool might request layout values or a list of assets. Better performance on Apple Silicon should make those interactions more responsive on compatible Mac hardware, although the announcement does not provide benchmark results. The OpenAI-compatible updates are useful for developers building around a common client interface. Tool search can help a client find or select tools during an interaction. Response compaction is intended to reduce the amount of conversation context that must remain active as a session grows. Together, these additions matter when a project involves repeated model calls, external actions, or long-running sessions. They may also make it easier to test an Ollama-backed workflow without rewriting every integration around a different API shape. Image support is another practical detail. The release notes say images now work correctly through compacted responses. That matters for applications that combine visual inputs with longer conversations. For example, a maker could keep an image review session going while the system compresses earlier context, rather than treating image handling as a separate path that may break when the response is compacted. On Mina Labs, Ollama v0.34.0 is available for 8 per image. That gives makers a direct way to try the release from the platform rather than setting up a local environment first. The listing is the useful starting point if you want to evaluate the model behavior, test a prompt flow, or compare the output with an existing project. The price token refers to one image; the release itself includes broader text, tool, and client-integration changes. Minaxlab would use this release first for structured creative briefs. A prompt could request a fixed object containing a project concept, target audience, visual direction, shot list, and production notes. The resulting fields could then feed a second step in a workflow. The performance improvements on Apple Silicon would be useful when iterating over many versions of the same brief locally. Minaxlab would also test it for image-aware review sessions and tool-driven production tasks. An image could be submitted for analysis, followed by requests for alternate compositions, captions, or asset metadata. Tool search and response compaction could support a longer session where the model needs to find an available action and retain the important decisions from earlier turns. These are practical tests for the release because they exercise the specific areas Ollama has changed: structured output, compatible clients, compacted responses, and images.