dalao.ai

Ollama

Get up and running with local modelsOpen source

About Ollama

Ollama is a tool for running open models on a local machine. One command pulls and starts a model; a local server on port 11434 speaks its own REST API plus OpenAI- and Anthropic-compatible formats, so most chat apps and coding agents can plug straight in. The code is MIT-licensed Go, built on llama.cpp and, on Apple Silicon, MLX. Cloud models arrived in preview in September 2025 and moved to per-token billing on 2026-08-31; the company announced an $88M raise on 2026-07-09. As of 2026-10-25: 184,664 GitHub stars, latest stable v0.40.11 (2026-10-24).

EditorialUpdated 10 min read

What it does

Open weights are free to download. Getting them to run is the hard part: the choice of quantization, choosing an inference engine, sorting out GPU drivers, getting chat templates and stop tokens right, and then exposing the model to other software. The founders tell the same origin story in their July 2026 funding post. The models were out there, but developers couldn't use them the way they used a proprietary model behind an API.

Ollama treats models like packages: they are pulled, run, listed and removed. A background service loads weights into memory on demand and unloads them when idle. By default it binds only to 127.0.0.1:11434 (per the FAQ), and other programs talk to it over HTTP. The README's first example is ollama run gemma4. The second is a curl call to /api/chat.

On compatibility, the docs list OpenAI-style endpoints for chat completions, completions, models, embeddings, Responses (stateless only) and audio speech. Since v0.14.0, released January 10, 2026, it also accepts the Anthropic Messages format at /v1/messages, which lets Claude Code and similar tools run against open models. The v0.40.9 notes fixed a case where streaming Responses and Messages requests failed silently.

The engine underneath has changed more than once. The README credits llama.cpp, created by Georgi Gerganov, as the supported backend. Version 0.30 (June 5, 2026) deepened GGUF support and turned on Vulkan by default, so AMD and Intel GPUs get acceleration out of the box. On Apple Silicon, an MLX-based engine went into preview on March 30, 2026. Starting with v0.40.0 (September 25, 2026), any model architecture MLX supports runs on MLX by default. The v0.40.2 notes warn that older downloads get upgraded in the background for the llama.cpp runner, with the original kept as a backup for now, which temporarily takes extra disk space.

Repo snapshot, 2026-10-25: created 2023-06-27, Go, MIT, 184,664 stars, 18,546 forks, and 4,410 open issues plus PRs. GitHub lists 269 releases. Four are flagged as prereleases, the most recent being v0.32.10 on 2026-08-13; everything else is a stable release.

Start building with open models. Access the latest releases locally or in the cloud with complete privacy.

Ollama README

Key features

A model library addressed by name. Anything on ollama.com/library can be pulled directly, in several sizes and quantizations. A GGUF from Hugging Face also works: the 0.30 post shows a one-line Modelfile pointing FROM at the file, followed by ollama create. If the source model supports tool calling, that carries over, and ollama show will list it.

Acceleration across hardware. Ollama says 0.30 (2026-06-05) runs up to 20% faster on NVIDIA. That's a vendor benchmark: Gemma 4 26B at Q4_K_M on an RTX 5090. The same release enabled Vulkan for AMD and Intel. The MLX preview post (2026-03-30) has its own vendor numbers, measured March 29, 2026. Running Qwen3.5-35B-A3B, prefill went from 1,154 to 1,810 tokens per second and decode from 58 to 112, comparing 0.18 at Q4_K_M with 0.19 at NVFP4.

One interface for local and cloud. Cloud models arrived in preview on September 19, 2025 (v0.12). They behave like local ones (run, pull, cp) but execute on ollama.com and require sign-in. Ollama's stated policy is that prompts and responses are never logged or trained on, that hosting is primarily in the US with overflow to Europe, and that its NVIDIA Cloud Partner hosts must commit to no logging, no training and zero data retention.

Coding agents in one step. ollama launch shipped on January 23, 2026 (v0.15+). It configures and starts Claude Code, OpenCode, Codex or Droid without touching environment variables. Ollama's docs give at least 64,000 tokens as the context length for coding tools. The FAQ puts the default at 4,096.

New capabilities. v0.40.11, released 2026-10-24 08:30 (UTC+8), added speech generation on macOS: ollama run can read text or a model's reply aloud as it streams, and apps can pull audio from /v1/audio/speech. The same release added d3 decision models for classification, yes/no questions and scoring. Earlier October releases brought NVIDIA Nemotron 3.5 and Recursive's Flint coding model.

How it compares

llama.cpp (MIT, 132,256 stars on 2026-10-25) is one of Ollama's backends; the README says so. It provides the engine, CLI and server with flags exposed directly. Ollama adds the registry, caching, a resident service, template handling and compatible APIs; new llama.cpp features arrive only after Ollama adopts them, and some knobs stay hidden.

vLLM (Apache-2.0, 95,347 stars) is a datacenter inference server built for throughput across many GPUs and many concurrent users. Ollama's default setting is one machine and one user.

LM Studio comes from Element Labs. Its desktop app is governed by the company's terms of service (version of August 23, 2026), which grant a non-exclusive, non-transferable license to use it rather than an open-source license, and it also runs models locally via llama.cpp and MLX. LM Studio's pricing page (as of 2026-10-25) lists Free at $0 for local models, then Bionic+ at $20/month and Pro at $100/month for US-hosted open-model cloud inference. Ollama is likewise free locally, with cloud Pro at $20 and Max at $100. LM Studio centers on a GUI and its Bionic agent, while Ollama centers on the CLI and API and is fully open source.

Against closed-model APIs, local inference has no per-token bill and data stays on the machine, but capacity is capped by local hardware, and large models need a lot of VRAM. Ollama's cloud runs models too big for a laptop, metered per token.

Who uses it, and who it's for

Adoption figures. The headline numbers come from Ollama. The July 9, 2026 funding post claims 8.9 million developers, use at 85% of the Fortune 500, and cloud token volume that has more than doubled month over month. Two public indicators are available. First, the README's list of community integrations: Open WebUI, LibreChat, Lobe Chat, Dify, AnythingLLM and Cherry Studio all treat Ollama as a default local model source. Second, downloads of the ollama Python client: as of 2026-10-25 09:49 (UTC+8), pypistats puts them at 13,275,724 for the past month and 2,786,399 for the past week, counted in UTC through 2026-10-24.

On contributors, the GitHub API (2026-10-25) shows 446 linked accounts. Names that keep turning up in recent release notes include jmorganca, dhiltgen, hoyyeva, jberg5 and drifkin.

Who it's for

It fits developers running an open model on a Mac, Windows or Linux laptop with as little setup as possible. It fits teams adding local inference to an app, since moving an existing OpenAI client over is mostly a base_url change, which matters when data can't leave the building. It fits people running Claude Code, Codex or OpenCode on open models; ollama launch wires that up in one command. And it fits anyone using the same CLI and API for small models at home and big ones in Ollama's cloud.

Not for

High-concurrency production serving on multi-GPU servers, where a throughput-oriented server like vLLM is the better fit. Tinkerers who need every inference flag and the newest quantization formats on day one; raw llama.cpp exposes more control. And any multi-user setup that would put Ollama directly on a network: it has no built-in authentication, and unauthenticated endpoints such as /api/create and /api/push keep showing up in advisories.

Getting started

Installation. Per the README, macOS and Linux use curl -fsSL https://ollama.com/install.sh | sh, and Windows uses irm https://ollama.com/install.ps1 | iex in PowerShell. Installers are also available for download, and servers can use the official ollama/ollama image on Docker Hub.

Running a model. Typing ollama on its own offers to start a model or connect Ollama to Claude Code, Codex, OpenClaw and other tools. Plain chat is ollama run gemma4; other model names are listed at ollama.com/library.

Connecting code. The server listens on localhost:11434. The native endpoints are /api/chat, /api/generate and friends. An existing OpenAI client just needs base_url set to http://localhost:11434/v1/. Official libraries are pip install ollama for Python and npm i ollama for JavaScript.

Tuning. Coding and long documents need a longer context than the FAQ default of 4,096 tokens, adjustable with OLLAMA_CONTEXT_LENGTH or in the app settings. K/V cache quantization reduces memory use. The FAQ rates the precision loss as very small for q8_0 and small-to-medium for q4_0, more noticeable at longer contexts. OLLAMA_HOST=0.0.0.0:11434 opens the server to other machines, and with it the unauthenticated endpoints.

The cloud is where bigger models run. As of 2026-10-31 the $0 tier on the pricing page is Pay as you go: no starter allowance, every cloud model is paid per token, with up to 3 concurrent requests.

Pricing & license

MIT, copyright "Ollama", so commercial use and redistribution are fine. Running models on local hardware costs nothing, and the pricing page calls local use "always unlimited." Cloud plans as of 2026-10-31:

Pay as you go: $0, with no subscription and no service fees; every cloud model is paid per token. 3 concurrent requests.

Pro: $20/month or $200/year, including $60 of usage a month. 3 concurrent requests.

Max: $100/month, including $300 of usage. 10 concurrent requests.

Team: early access at $500/month, with unlimited users sharing $1,000 of monthly usage.

Enterprise: custom pricing.

Models are priced per million tokens, and the pricing page gives off-peak rates outside weekday peak hours and all day at weekends, counted in UTC; converted, peak hours are 20:00–02:00 (UTC+8) each weekday, running past midnight. This per-token scheme launched on August 31, 2026 and removed the old five-hour and weekly caps. Existing subscribers can stay on legacy plans as long as they don't change them. Each model also carries its own license, separate from Ollama's.

Risks & limits

On security, the ollama/ollama repo has published no GitHub security advisories of its own (zero via the API on 2026-10-25), but the GitHub Advisory Database lists 16 reviewed advisories affecting Ollama. The three most recent are as follows.

CVE-2026-7482 (GHSA-x8qc-fggm-mpqg, added 2026-05-04). GitHub rates it high, while the same page shows a CVSS 3.1 score of 9.1 and a CVSS 4.0 score of 8.8. The advisory describes a heap out-of-bounds read in the GGUF loader in versions before 0.17.1. /api/create accepts an attacker-supplied GGUF whose declared tensor offset and size run past the end of the file, so quantization reads beyond the heap buffer. Leaked memory can include environment variables, API keys, system prompts and other users' conversations, and it can be exfiltrated through /api/push to an attacker's registry. The advisory notes that both endpoints are unauthenticated upstream. Default installs bind to 127.0.0.1, but the documented OLLAMA_HOST=0.0.0.0 setting is common, and there is large public exposure. NVD (status Analyzed) shows only the CNA's scores, 9.1 CRITICAL under CVSS 3.1 and 8.8 HIGH under 4.0, with no separate NVD score. The fix is in 0.17.1.

CVE-2025-63389 (GHSA-f6mr-38g8-39rg, added 2025-12-19). GitHub rates it critical, with no CVSS 3.1 score and a CVSS 4.0 score of 9.3. The advisory text says multiple API endpoints in v0.12.3 and earlier require no authentication, letting remote attackers manage models. GitHub's affected range is 0.13.5 and earlier, and no patched version is listed. NVD shows only a CISA-ADP CVSS 3.1 score of 9.8 CRITICAL, with no NVD score of its own.

CVE-2026-7020 (GHSA-x99g-8v8j-25j2, added 2026-04-26). GitHub rates it low, yet its CVSS 3.1 score is 5.6 (CVSS 4.0: 2.9). The advisory describes a path traversal in digestToPath in x/imagegen/transfer/transfer.go in versions up to 0.20.2. It can be triggered remotely but is high-complexity and hard to exploit. A public exploit exists, and the advisory says the vendor never responded to early contact. NVD shows the CNA's (VulDB) scores: 3.7 LOW under CVSS 3.1, 2.9 under 4.0 and 2.6 under 2.0, with no NVD score of its own. No patched version is listed.

Ollama's API has no authentication, and its safety rests on listening only on localhost. Once it listens on 0.0.0.0, Ollama itself provides no access control; authentication and network isolation have to be added separately. Other limits: the default context is 4,096 tokens. Releases are frequent, so unpinned production installs change with them. Since v0.40.2, older models are upgraded in the background with backups kept, which temporarily costs disk. User and Fortune 500 figures are self-reported.

Our verdict

What sets Ollama apart: one command pulls and runs an open model; it speaks the OpenAI and Anthropic formats, and clients in the README's community integrations, such as Open WebUI, LibreChat and Dify, treat it as a local model source; the local runtime is free and MIT-licensed, the cloud is billed per token, and the company announced an $88M raise on 2026-07-09.

Its limits: the API has no authentication of its own, so network-facing deployments need separately added authentication and isolation and timely upgrades; the default context is 4,096 tokens; user and Fortune 500 figures are self-reported.

Why it stands out

Ollama gave local models a package manager and a one-line install. `ollama run` does for weights what `docker run` did for containers: fetch, cache, serve, and expose a clean API on localhost.

Jeffrey Morgan and Michael Chiang came from Docker, and the influence is deliberate. Modelfiles, a registry, a daemon that just works on a Mac, and an HTTP interface that every editor and chat client now speaks.

One command fetches, caches and serves a model, behind an HTTP API on localhost that editors and chat clients use.

Sources

Official

  1. ollama/ollama (GitHub)GitHub · Accessed Oct 25, 2026
  2. Ollama READMEGitHub · Accessed Oct 25, 2026
  3. Ollama LICENSE (MIT)GitHub · Accessed Oct 25, 2026
  4. ollama/ollama ReleasesGitHub · Accessed Oct 25, 2026
  5. Release v0.40.11 · ollama/ollamaGitHub
  6. Release v0.14.0 · ollama/ollamaGitHub
  7. Release v0.40.0 · ollama/ollamaGitHub
  8. Release v0.40.2 · ollama/ollamaGitHub

Updates & reading