Ollama Review

Ollama is the runtime many developers install first, and in 2026 that install now sits next to a cloud price list. The download is still the right call when the weights have to stay on the machine, since that path has no vendor bill. The paid cloud plans are a poor default for a small model, because pay-as-you-go on the free account can undercut the subscription.

Devshot data: Toolradar, the software directory we run, tracks 45 AI model deployment tools as of 17 October 2026. On its deployment ranking, 4 of the 10 listed tools publish a paid price, and the median cheapest paid plan is $64.50 a month. Ollama's local runtime is not in that paid median, so the free install sits below that middle. The cloud tiers are the bill this review works through.

Below, the credit pools, the concurrency caps and a token month priced from the published rates. Hosted APIs with a similar job are in our LLM API roundup and the free LLM APIs note. A wider desktop list is on Techpresso's local LLM tools guide. Toolradar profiles for Ollama, LM Studio and vLLM sit beside this comparison.

Ollama is a local runtime and a cloud API for open-weight models. It gives you a desktop app, a CLI and an OpenAI-compatible endpoint, and the same account can run a model on your GPU or on Ollama's servers. The software in the GitHub LICENSE file is MIT, so you may use, modify and ship it. Running models on your own hardware is unlimited, while cloud tokens are metered and can become a bill.

Installers exist for macOS, Windows and Linux, and the official Docker image is ollama/ollama. NVIDIA support starts at compute capability 5.0 with driver 550 or newer, and cards in the 5.0 to 6.2 range need driver 570 or newer. A card or driver below that floor closes the free local path.

AMD GPUs go through ROCm, and Linux needs the ROCm v7 driver. An older AMD stack has the same limit until that driver is current.

Cloud prompts are a different product from the MIT binary. Ollama says prompt and response data is never logged or used for training, that compute is hosted primarily in the United States, and that traffic may be routed to Europe for capacity. Partners are NVIDIA Cloud Providers, and Ollama says it requires those partners to keep a zero data retention policy. If the machine is the whole point, the cloud docs say to disable cloud features.

You install the app or the script, then run a model locally or sign in for cloud models. Local OpenAI clients point at http://localhost:11434/v1/, where the client library insists on an API key and the local server ignores it. Cloud clients point at https://ollama.com/v1 with a real key, and they do not need the app installed. The cloud endpoint is a subset of the OpenAI API: no stateful Responses, no built-in web search through /v1/responses, and no custom tool-call replay.

An agent that needs those calls will not run on this hosted path.

Included credits are spent before credits you buy, so a quiet month burns the paid pool first. Paid pools reset on the subscription date, including on annual plans, and unused credits expire, so prepaying the year does not bank a larger pool. The free allowance resets from signup, and its dollar size is not printed, so you cannot forecast that free month from the pricing page. Buying credits on Free unlocks every cloud model with no service fee, and paid plans email you at 90 percent of included usage.

Concurrency is the cap on how many agents can run at once. The free plan is the smallest cap, and the paid tiers raise it. Extra requests queue, and a full queue rejects the call until a slot opens, which fails the turn rather than slowing it.

For coding agents, the docs ship launch commands for Claude Code, Codex and Copilot CLI, plus OpenCode. Our Python coding tools note is the place those agents get compared on the editor side. ollama ps shows whether a local model landed on GPU, CPU or a split, which is the check to run before you trust a latency number.

Prices checked on 18 October 2026 on Ollama's pricing page. Local use stays free, and it is the whole product when the weights stay on your machine. The free cloud plan is $0, with starter credits for a smaller model set, one concurrent request, and no published dollar size for those starter credits. The missing dollar figure means the free month cannot be forecast before signup.

The annual Pro row is $200/yr, which works out to $16.67 a month, and included usage still resets monthly. The discount is on the fee, not on a larger credit bank.

Pro adds $60 of credits, larger models and 3 concurrent requests. A fast mode is marked coming soon, so it is not part of what you buy today.

Max is the power-user tier: a larger credit pool, early access to new models and 10 concurrent requests. Team is $500/mo with $1,000 of shared credits, unlimited users and centralized billing. Ollama's 31 August 2026 post calls the Team price introductory. A long-term budget should not treat that sticker as fixed.

Enterprise is custom, and it adds model access controls, per-user and per-key budgets, and a private Slack channel. People already on an older Pro or Max stay there until they change tier or billing cycle, and new signups start on these token plans. The post says the previous bill was GPU time, which is why a 2.8 trillion parameter model such as Kimi K3 was hard to forecast.

Take kimi-k2.7-code at the peak rates in the table: $0.95 per million input tokens, $0.19 cached, and $4 per million output tokens. A month of 40 million uncached input tokens, 40 million cached input tokens and 8 million output tokens is $38 plus $7.60 plus $32, which is $77.60.

Usage past the included credits is $17.60, so the cash total on Pro is $37.60. The free plan's included usage covers a starter set, and buying credits unlocks every model at the same rates with no subscription. If this month is billed entirely to purchased credits, Free costs the token bill and Pro costs less. Max's pool covers the bill, so you pay only the Max subscription, which is more than Pro at this volume.

The same token counts on gpt-oss:20b, at $0.07 input, $0.035 cached and $0.30 output, come to $2.80 plus $1.40 plus $2.40, or $6.60. That sits under the Pro subscription, so Free plus purchased credits is the cheaper cloud path for a small model.

Pro costs less than Max until the token bill passes $140. Above the included credits, Pro's cash cost is the token bill minus $40, the gap between the fee and a credit pool three times as large. Max stays a flat subscription until its own pool is used up.

Five Max subscriptions match the Team sticker and, at five times Max's credit pool, hold $1,500 of credits that cannot move between people. Team is the better buy when one person burns the pool and others barely touch it. Separate Max plans are better when each person would have used their own credits.

Off-peak rates apply outside 12:00 to 18:00 UTC on weekdays and all day on weekends. A job that can wait outside that window pays less. deepseek-v4-pro is the worked example in the table: its off-peak input and output are half the standard rates.

LM Studio's pricing page lists a free local app, Bionic+ at the same monthly fee as Ollama Pro, and Pro at $100/mo. Pro says the usage limit is five times Bionic+, and neither paid row prints a credit-pool dollar amount. LM Studio's terms, version 23 August 2026, allow personal and internal business use and forbid offering the app as a service bureau or SaaS.

llama.cpp, vLLM and Jan publish no paid plan, so the GPU bill stays with you. LocalAI publishes no business list price, so a commercial deal cannot be compared here.

Methodology: Ollama figures come from the pricing page, the 31 August 2026 post, and the cloud, context-length, hardware and OpenAI compatibility docs, read on 18 October 2026. The license text is the GitHub LICENSE file from the same day. LM Studio's fee and terms are the pages linked above. llama.cpp is MIT on its LICENSE, vLLM is Apache 2.0 on its LICENSE, and Jan is Apache 2.0 on its LICENSE. LocalAI's MIT label and v4.8.0 were read on localai.io. No third-party price list was used. The token month is arithmetic on the published per-million rates, not a timed run.

Ollama fits a developer who wants an OpenAI-compatible server on the laptop and a cloud fallback with printed token rates. Solo builders who never overlap cloud calls, or who never turn cloud on, get the MIT runtime at no software fee. A coding agent that needs 64,000 tokens of context will want a larger GPU or the cloud, because the smaller-card default is too short and raising it spends VRAM.

It is a weaker fit when the company needs a single-region promise. Hosting is primarily in the United States and may move to Europe, which is not an EU-only or US-only commitment. A contract that names one country should stay on the local binary. It is also the wrong buy when you need to embed the desktop shell in a product you resell: LM Studio's terms block that, and Ollama's MIT license does not, but the cloud account is still one person per login.

A resold product cannot share one paid seat across its users. Teams comparing hosted routers should read OpenRouter alternatives and the coding assistant roundup before they add another seat.

Skip the cloud tiers when the job is a private repo on a machine you control. Our self-hosting tools and free AI coding tools roundups cover that path, and cheap GPU cloud is where the hardware bill lives if the laptop is too small. An editor with its own subscription, such as Cursor, is a different purchase from this runtime.