Ollama v0.34.1

Ollama is the local model runner from Ollama Inc., and this reading is of v0.34.1, released on 14 September 2026 and read two days later. It installs on macOS, Windows and Linux, pulls open-weight models onto the machine in front of you, and answers on a local HTTP API on port 11434. The code is MIT-licensed.

We read 18 sources: five official, being the README, the quickstart, the licence, the v0.34.1 release notes and the pricing page; three editorial reviews and how-tos; two trade-press reports; the repository's figures from the GitHub REST API; three developer threads; and four review sites outside the trade press. They are dated between March 2025 and 14 September 2026. No source we could reach wrote about v0.34.1 itself, so the press, editorial and independent reviews all describe earlier release lines, and only the official tier and the repository figures speak for the version named above.

Consensus

The agreement is on the shape of the product rather than on this week's build, and none of the eight press, editorial and independent reviews disputes it: a local runner that installs in one command, carries a large and quickly refreshed model library, and speaks an OpenAI-shaped API on localhost. The repository figures give the scale, 181,074 stars, 17,892 forks and 4,013 open issues on 16 September 2026: how many people looked by that date, and not how many are satisfied. The published terms: the local path is $0 with no service fee, Pro is $20 a month with $60 of usage credits, and Max is $100 with $300.

Confidence is uneven by tier. The official tier is strong, because the release notes, the docs and the licence describe v0.34.1 directly. Every other tier is thin for this version, and the editorial tier is thin twice over at three pieces covering earlier builds. We read the consensus as moderate on what Ollama is for and weak on how v0.34.1 behaves.

Recurring strengths

One command, and a model answers. Six of the sources we read describe install and first run as the part Ollama got right, two of them editorial reviews: the README and the quickstart give one line per platform, and the ZDNET pieces of February and May 2026 both describe a first response with no key and no configuration file.

A model library that keeps up. Five sources, one editorial and two official: the README points at a library rather than a fixed set of models, the v0.34.1 release notes record backend updates shipped alongside the release, and the March 2026 report describes a new Apple backend arriving inside an ordinary release.

A local API other software already speaks. Five sources, among them the official quickstart and an editorial how-to: the server answers on port 11434 in the OpenAI chat format, so a tool written for a hosted model reaches a local one by changing an address.

Free on your own machine, under a permissive licence. Four sources, three official and one editorial: the licence file is MIT, the pricing page lists the local path at $0 with no service fee, and the May 2026 ZDNET piece lists the fees it stopped paying.

The work stays where the model runs. Three sources, one of them editorial: the May 2026 ZDNET piece keeps sensitive files off another company's server, the August 2026 independent review puts control of the data first, and the August 2026 trade-press report notes the models can run on the machine rather than at a provider.

Recurring complaints

Memory sets the ceiling. Five sources across three tiers describe it. The February 2026 hands-on piece ran small models on a 16 GB M1 MacBook and advised 32 GB. The May 2026 piece puts the floor at 16 GB of memory, with 8 GB of video memory or Apple Silicon with 16 GB or more. The March 2026 report gives 32 GB as the requirement for the Apple backend preview, and two independent reviews put 16 GB as the minimum and 32 GB as the practical baseline. Neither the README nor the quickstart states a minimum amount of memory. No source we read counts the users who hit that wall, and a count of complaints is not a failure rate, so the gauge beside this complaint reads unknown.

Concurrency is not what it was built for. Two independent reviews describe four parallel requests by default, no continuous batching and no paged cache, and say throughput flattens once real traffic arrives.

The default context window is easy to get wrong. Two sources. The August 2026 review describes a 4,096-token default that drops the front of the prompt past it without saying so. The March 2025 repository issue found a model advertising 131,072 tokens running at 2,048 until the size was set.

A release can change behaviour under you. Three sources: the August 2026 report describes a Claude Desktop integration that lasted days, was withdrawn and returned as a proxy in v0.33.0; the March 2026 report describes the Apple backend arriving as a preview for one model; and the v0.34.1 notes record a setting deprecated.

The API is unauthenticated where it listens. One source, the Go vulnerability database listing for this module, records missing authentication that lets an attacker perform model management operations, under CVE-2025-63389 and GHSA-f6mr-38g8-39rg, published 14 January 2026. We state that as the database states it.

Where reviewers split

Whether the layer should exist at all. On one side, the April 2026 Hacker News thread on a piece titled The local LLM ecosystem does not need Ollama, with 207 comments, argues that a wrapper between you and the model is unnecessary. On the other, the May 2026 ZDNET piece is written by someone who stopped using a hosted assistant for the same work. Our read: the thread is not unanimous, and its most upvoted reply credits the project with solving the first run.

How much context you get by default. One side, the August 2026 review, describes a 4,096-token default and a prompt front dropped past it without warning. The other, a repository issue from March 2025, found a model advertising 131,072 tokens running at 2,048. Our read: the default has moved between those dates, and both describe a limit that is silent rather than fixed.

Who it suits

You want open-weight models on a machine you already own, with the weights and the prompts staying there. The official pages and two independent reviews describe that setup, and the local path is $0.

You write against an OpenAI-shaped API and want the same call to reach a local model. The README, the quickstart and the August 2026 review all describe one call working at both ends.

You are one person, or a few, prototyping against a model you host. Both independent reviews that measure traffic describe a small-team load.

Who should pass

You are counting on a 16 GB laptop for anything beyond the smallest models. The February 2026 hands-on piece advises 32 GB, and the independent reviews put 16 GB as the floor.

You need throughput behind one endpoint for a team. Both independent reviews that measure traffic describe a ceiling on parallel requests.

You are about to listen on an address other machines can reach. The vulnerability listing of 14 January 2026 records missing authentication on model management operations; we report that listing, not a verdict.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it. Send me the weekly verdict

Sources

• Ollama README, official, read 16 September 2026.
• Ollama quickstart, official, read 16 September 2026.
• Ollama licence file, official, read 16 September 2026.
• v0.34.1 release notes, official, 14 September 2026.
• Ollama pricing, official, read 16 September 2026.
• ZDNET, I tested local AI on my M1 Mac, editorial, 1 February 2026.
• ZDNET, I quit ChatGPT for Ollama, editorial, 26 May 2026.
• The New Stack, Connect to a Local Ollama Instance on Your LAN, editorial, 6 September 2025.
• The New Stack, Ollama taps Apple's MLX framework, news, 31 March 2026.
• The New Stack, Claude Desktop can run Qwen, DeepSeek and Kimi, news, 26 August 2026.
• GitHub REST API, ollama/ollama figures, tier 3, read 16 September 2026.
• Hacker News, The local LLM ecosystem doesn't need Ollama, forum, 16 April 2026.
• Hacker News, Ollama powered by MLX on Apple Silicon in preview, forum, 31 March 2026.
• GitHub issue 9890, Large context size breaks usability, forum, 19 March 2025.
• aifoss.dev, Ollama 2026 Review, other, 14 May 2026.
• Francis Okafor, Ollama Review: Pricing, Pros and Cons, other, 28 August 2026.
• AIwithKay, Ollama Review, other, 14 August 2026.
• Go vulnerability database, GO-2025-4251, other, 14 January 2026.

This reading covers 18 sources dated between March 2025 and 14 September 2026 and read on 16 September 2026, across official documentation, trade press, independent reviews, the repository's own figures and developer threads, synthesised with AI; nothing was installed, prompted, run or tested.