Hammer AI Review 2026: the model runs on your own machine

Local or cloud: what privacy at Hammer AI really covers

The FAQ answers the question of whether the chat is private with a sentence I never otherwise read in this category: it depends which models you use. Then it draws the line itself. In the desktop app with local models the conversation does not leave the device. With cloud models it goes to the provider’s servers, where by the provider’s own account it is neither stored nor used for training. That is the clearest separation I have met in two years of testing, and it belongs in front of every criticism.

What stays on the machine in local mode

Everything: the model file, the character as JSON, the chat history and the lorebook sit in a folder on your drive, and the docs say which one. The install guide explains that Hammer AI opens a local server on a random port to talk to Ollama, and that the firewall prompt on first launch only needs internal access. The desktop page adds that messages are deleted on close as long as you do not save the history. I checked the only part a user can check: in the middle of the interrogation scene I cut the connection and kept playing. The character held the thread as if nothing had happened.

Two caveats belong with that. First, the desktop app checks for updates and downloads them automatically, so it is not a sealed box. Second, the source code is closed and shipped through a GitHub repository that holds releases only. So “nothing leaves the device” is documented, observable behavior, not audited behavior. That is enough for me, because I can reproduce the behavior myself; anyone who needs third-party verification will not get it here.

What the cloud side collects on top

In the cloud Hammer AI is an ordinary web service, and the privacy policy dated March 29, 2026 reads like one. The good sentence first: no information is collected about your conversations with characters. After that comes what very much is collected, namely the characters you create, usage statistics through a self-hosted Umami instance, payment data through Gank, feedback including your IP address, and error reports through Sentry. Two ad networks are named as well, even though the pricing page promises an ad-free experience on every plan, which is why I assume they cover the marketing pages and not the chat.

The line I would underline concerns the model providers: OpenRouter is named as a provider, and messages sent to third-party AI providers can be stored there under their rules. So “not stored” is a statement about the provider’s own servers, not about the whole chain. Since version 1.0.0 in March 2026 there has been a login with cloud storage for characters, which is the only reason the phone app works at all. The trade is simple: sign in and your characters follow you onto the phone; stay signed out and they stay in a folder on one device. The account is offered to you with striking restraint, and that restraint is the most trustworthy thing about the whole app.

What hardware Hammer AI needs and which model to pick

The home page promises that everything runs with no configuration. That is true of the setup and says nothing about the speed. Neither the home page nor the desktop page names a minimum spec, and I consider that the biggest mistake in the communication, because that exact question decides whether the evening is great or unbearable.

What numbers the docs give you

The honest figures are in the quantization docs. At Q4 a 3B model is a file of 2 to 6 GB and wants roughly 6 to 10 GB of memory. A 7B is 4 to 8 GB on disk and about 6 to 10 GB of VRAM. A 13B is 10 to 12 GB on disk and about 12 to 14 GB of VRAM. A 30B lands at a 24 to 30 GB file and 26 to 36 GB of memory, which rules out every consumer card. Q8 roughly doubles those numbers. The Ollama docs add that a model too large gets split between graphics card and processor, which works and is slow, and that an SSD helps because paging to a spinning disk slows everything further. The Linux docs name the symptom outright: slow or stuttering replies mean the model needs more memory than is there. For AMD cards in the 9000 series there is an extra restriction to Windows and specific Ubuntu and RHEL versions. On Apple Silicon the shared memory counts, which is why a MacBook with 16 GB runs a 12B where a Windows laptop with integrated graphics crawls.

How I picked the model

Take the biggest model that fits into your VRAM at q4_K_M, then test whether the next size up is still bearable. That is the docs’ advice and it held up for me. The HammerAI account on ollama.com is the practical shopping list: up front sit a 12B merge with over 40,000 downloads, MythoMax 13B, a 7B for short dialogue, a 24B for people with a big card, an 8B Lexi Uncensored and a 70B for the obsessed; the collection runs from 7B to 123B. Independent tests this year name Mistral-Nemo 12B as the free cloud model with a paid model above it, with the expected trade of better prose for a longer wait.

In the interrogation scene the difference was clearest. On the 7B my customs officer started mirroring my own phrasing after three turns. On the 12B she used the same card to ask me a counter-question I was not ready for. If none of the models suit you, the proxy setting takes your own endpoint with an address, model name, context length and key; that turns Hammer AI into a front end for whatever you already pay for. The browser build also runs locally through WebGPU in Chrome, Edge or Opera, but in practice it reloads the model every session and is slower than the desktop app. Treat it as a test drive, not a product.

Hammer AI in roleplay: character cards, lorebooks and group scenes

The character system is the familiar card model with a friendlier face: a card with personality, scenario, first message and example dialogue, a lorebook for facts, a persona for yourself, and a public library where the community parks its characters. The library sorts by popularity, has a featured filter and over fifty tags. No official page names a total character count; the App Store text says thousands.

What the form wants from you

The creation form asks for a name, an optional display name, personality as bullet points or prose, scenario, first message, example dialogue in a fixed format, avatar and background, an intro line, tags, a creator note the model never sees, and a rating of Safe or Mature. On top of that every character carries its own system prompt and its own sampler values, and that is exactly where the app pulls ahead of hosted services: a character who needs a low temperature to stay in her voice gets it, with nothing changing for anyone else. Save locally and the character stays private. Upload and it goes through a review that the docs say takes hours to a week, with mandatory tags and the content guidelines: all characters must be adults, avatars stay work-safe, photorealistic real people are banned. Which fields in which order actually pay off is written out at length in my guide to building your own AI character.

How a lorebook really fires here

A lorebook entry is a set of keywords and a value. If a keyword shows up in the last four messages, two from you and two from the character, the value is pushed into the prompt once. The docs are precise about the rules: case does not matter, partial matches count, duplicate hits push only once, and a lorebook value never triggers another entry. The advice to cut values down to one or two sentences and keep keywords sharp is right: a vague keyword in a long lorebook will flood your context.

What matters is what a lorebook is not, namely a memory. It remembers what you wrote into it, not what happened. For my campaign I split it like this: places, factions and names into the lorebook, the state of the plot as three lines in the first message, refreshed after every session. Group chats take up to ten characters, and they are as chaotic as group chats always are; two characters with opposing goals was the maximum where I could still steer a scene. How I build a campaign that survives a week is written up step by step in my AI roleplay chat guide.

The Hammer AI memory test after seven days

On day seven I carried the same storyline forward in three variants: locally on the 12B, on the free cloud model with 8,192 tokens, and in Starter with 16,384. The result is unspectacular and entirely logical. In the free context the scene had lost its premise; the character knew my name but not the reason we were sitting in that harbor at all. In Starter the premise was still there, but three side characters had vanished. Locally with the context set high it held the most, and I paid for that in sluggish replies.

The pricing page lists automatic memory on every plan, but the docs describe no summarization system, and the lorebook only fires on the last four messages. In practice that means: what is not in the window and not hit by a keyword does not exist. That is why memory sits at 6 here, even though the tools are better than at most hosted apps. Locally you get to set the context size yourself, which is a real advantage, but a 12B at 32,768 tokens wants memory most laptops do not have.

My quiet evening showed it most clearly. In that scene nothing happens on purpose: two people, one table, a long evening with no plot. A character who has only the context window fills that emptiness with questions about me, because nothing else is left. A character with a maintained lorebook fills it with her own past, and suddenly the scene works. At Hammer AI the difference was not the model, it was three lines I had written myself on day three. That is the whole mechanism of this app in one sentence: it hands you the tools and takes no notes for you.

The working way to deal with it is the same as everywhere: keep a short state of play outside the chat and refresh it. Anyone who does not want to is better served by an app that does the remembering itself; Nomi AI is the strongest choice for that in my tests and costs accordingly. For comparing free tiers across the whole category I built a separate list of free AI chat apps, where Hammer AI sits near the top for one reason: the free offer here is not an appetizer, it is the product.

What Hammer AI really costs per month

The short answer: nothing, if your machine can carry the model. The long answer has three line items, and only two of them are on the pricing page.

First, the subscription. Starter costs $9 a month, Advanced $18, Ultimate $35, none with an annual discount, because there is none. In the iOS app the monthly prices sit one cent lower, but there you also get weekly purchases that work out more expensive per month: $2.49 a week is substantially more over a year than $9 a month. Second, the credits. Every image pulls 10 credits, every voice clip 5, and Starter delivers 1,500 credits a month, so 150 images or 300 clips. Subscription credits expire monthly, purchased packs do not, and the packs run from $1.99 for 200 credits to $99.99 for 20,000.

Third, the hardware, and that is the item nobody writes about. If you already have a card with 12 GB in the machine, you effectively pay nothing. If you would have to buy one, you are comparing a three-figure one-off against $9 a month and should be honest about how many evenings a year you really play. My own case after two weeks: local on the desktop is enough for me, and I canceled the Starter subscription after the test month, because I do not write roleplay on a phone anyway.

Two more things about the money. Web payments run through Gank, and the pricing page says explicitly that all purchases are final and refunds are not supported; App Store subscriptions you cancel in your device settings. And the pricing page has already contradicted itself on credit and model counts this year, which is why independent checks and the page occasionally disagree. Trust the number at checkout, not the one on the marketing page.

Support, updates and the rough edges of a one-man project

The community is a Discord server with around 5,000 members, a subreddit, and a comment page on itch.io that effectively serves as a public bug tracker. Support is that same Discord plus an email address. There is no ticket channel, no live chat, and two different contact addresses across the docs and the policy, which tells you pretty precisely how big the shop is. The docs name a channel for character suggestions and a public roadmap channel where planned features are visible, and they say openly that much of it is prioritized by what people ask for.

The itch.io comments are the most honest source about this app. Over three years they run from “when is a Windows version coming” through crashes and complaints about too little VRAM to the protest when the subscription arrived in 2024. They are worth reading alongside the fact that local chat has stayed free the entire time. The developer answers in most threads, sends people to Discord and explains the funding without ceremony: the money coming in pays for more developers. By mid-2026 over 148 releases had accumulated, which is a remarkable cadence for a project this size.

On top of that comes a quirk that cost me two evenings: updates. If an auto-update hangs, the docs advise leaving the app open for five to ten minutes or pulling the current build straight from the release repository on GitHub. Both work, and neither is a procedure you hand to a casual user. Anyone starting out here should bookmark the GitHub page and know that an antivirus tool may complain about the installation, because the signature comes from a private source.

The ratings reflect both sides. The desktop listing on itch.io sits at 3.7 out of 5 from 25 ratings, the iOS app at 3.2 from 17. That is not enthusiasm, and it is not a wreck either; it is the signature of a project that ships fast and leaves edges. If you need a provider with a support department, Hammer AI is the wrong choice, and my list of the best AI character apps names the hosted alternatives in the order I would recommend them. If what matters more to you is that the model runs on your own desk, there is currently no second app that offers it this plainly.