This page compares eight offline AI drives. Seven of them you can buy and use today; one is not shipping yet. The marketing reads almost identically across all of them: an AI that works with no internet, a library of survival manuals, no subscription. Almost all of what follows is built from what each company publishes about its own product. There is one exception, a drive we could not get to run, and it is flagged as such at the end.
One rule shapes everything below. Where a company does not mention a capability, this page says “not mentioned on their site” rather than claiming they lack it. That is what was actually checked, a search of their pages returning nothing, and it is not the same as the feature being absent.
A note on names. OffGrid AI Toolkit, at offgridaitoolkit.com, is a separate company with no connection to Off-Grid AI. The two names are close enough to be confused, so every product below is listed with its web address.
Every competitor claim read from their own website, 21 and 22 August 2026
The comparison
Eight products across nine columns. The table scrolls sideways, and the row labels on the left stay in place while it does.
| Off-Grid AI offgridai.io |
Docket Survival docketoffline.ai |
Trove trovedrive.com |
DrAcher AI drachershop.com |
OffGrid Drive offgriddrive.com |
BunkerAI bunkerai.io |
PortableMind portablemind.io |
OffGrid AI Toolkit offgridaitoolkit.com |
|
|---|---|---|---|---|---|---|---|---|
| Price | $149.95 to $759.95 | $69.99 | $69 founder rate, $99 list | $119 to $199 depending on bundle | $49.95 | $70. Their own FAQ on that page says $89, and an older live page lists $89 | $69 download only, $99 with a drive, $139 faster drive, $199 PRO Navigator | $129, $249 or $469 |
| What arrives | A USB drive; capacity and tier vary | Ships a 64 GB USB drive | Not shipping yet. Pre-Kickstarter, availability listed as PreOrder | Ships a 128 GB USB drive | No hardware. A download; you supply the USB stick | Ships a 128 GB USB 3.2 flash drive in an aluminium shell | $69 is a download and you supply the drive. $99 and up ship a 64 GB USB drive. $199 adds a USB GPS receiver | Ships a 64 GB SanDisk USB 3.2 flash drive |
| AI models | Five ship, all named: Qwen3.6 35B-A3B, Qwen2.5 14B, Llama 3.1 8B, Phi-4-mini and Qwen2.5-VL 3B for vision, every one Q4_K_M quantised, plus a vision projector. You can add as many GGUF models as the drive has room for | States 10 pre-loaded models. No parameter sizes given | One model, named: Microsoft Phi-3.5, 4 billion parameters | States 13 models, named only as its own aliases. A 7B vision size appears in image text, not body copy | One model, "7B parameter", not named | States six models, named only as Fast, Reasoning, Logic, Math, Expert and Creative. No model names or sizes published | States "the Qwen 3.5 roster" with no count. Their benchmark page names a 4 billion and a 9 billion parameter model | Gemma 3 at 4, 12 and 27 billion parameters, plus MedGemma 4 billion. The frontier models in the $249 and $469 tiers run in the cloud and need internet |
| Reference corpus | About 1,100 documents totalling 161,286 pages across 18 knowledge domains, as of August 2026 and growing with each update. Six Expansion Packs sold separately | 30+ survival guides | 380+ hand-curated books and manuals, plus full English Wikipedia (about 90 GB) | 18,003 documents, sourced from FEMA, CDC, EPA and the US Army | 31 survival guides | A survival, medical and repair library. No document count published. Wikipedia not mentioned | A six category field manual: water, fire, first aid, shelter, signal, food. No document count published. Wikipedia not shipped; their free guide tells you to fetch it yourself | No document library. What ships is a prompt library, given as 500+ on one page and as hundreds on another |
| Offline maps | Streets, terrain, topographic contours and mapped trails, expandable with regional Map Packs | A trail map PDF among its documents. No mapping application mentioned on their site | Americas vector map. Pan, zoom, pins, notes, place search. Their site states it is "not a GPS tracker" | World vector map to US street level, with a distance ruler, 32,000+ town search and saved places | Not mentioned on their site | Not mentioned on their site | Only in the $199 PRO Navigator, Windows only. Full US road basemap with a stated 258 million address index | Not mentioned on their site |
| Topographic, terrain and trail data | Yes, where the installed Map Packs carry it | Not mentioned on their site | Not mentioned on their site | Not mentioned on their site | Not mentioned on their site | Not mentioned on their site | Not mentioned on their site | Not mentioned on their site |
| Offline route calculation | Yes | Not mentioned on their site | Not mentioned on their site | Shown inside an illustrative demo panel, not listed as a feature | Not mentioned on their site | Not mentioned on their site | Yes, in the $199 PRO Navigator. Turn by turn across the US road network | Not mentioned on their site |
| Answers cite their source | Yes. Every Intel answer names the document it came from | Not claimed. Says answers pull from the loaded manuals | Not claimed. Describes grounding in reference material | Yes. States answers carry [1] citations to the library | Not claimed. Describes RAG retrieval | Not claimed | Not claimed | Not claimed |
| Declines when evidence is thin | Yes in Intel, the cited path and the default. Companion is open chat and is not cited; its Unrestricted toggle removes our instructions to the model, but not the refusals trained into the model itself | Not claimed. Markets "uncensored" models | Not claimed. Publishes a caution that it is "not a final authority" | Not claimed. Markets "unrestricted" models | Not mentioned on their site | Not claimed. Markets "uncensored" models and sells a $40 jailbreak upgrade | Partly. Photo identifications carry a confidence rating, and their own published benchmark grades 9% of survival answers wrong on the small model. They also sell an "Uncensored Pack" | Claimed. A prompt layer instructs the model to flag claims it cannot support. Self graded, no independent audit |
| Photo identification | Yes (Scout), and it declines edibility judgements by design | Yes. Plants, terrain, injuries | Not mentioned on their site | Yes. Plants, mushrooms, gear, wounds, plus OCR and translation | Not mentioned on their site | Not mentioned on their site | Yes. Seven scan modes, each returning a confidence rating and dangerous lookalikes | Yes, on the Gemma 3 models. Their own docs note MedGemma does not read images |
| Stated hardware requirements | 8 GB RAM minimum, 16 GB recommended. An Intel Core i5/i7 or AMD Ryzen 5/7 from 2020 or later. A GPU is optional but cuts response time sharply: NVIDIA on Windows, Apple Silicon on macOS 14+. The host computer does the work, not the drive. Why your computer matters more than the drive | Vague: "any modern computer with a USB 3.0+ port" | 8 GB RAM minimum, 16 GB recommended, no GPU required | None stated anywhere on their site | 8 GB RAM, macOS 11+ or Windows 10/11 64-bit | 8 GB RAM recommended. No CPU or GPU requirement published | 16 GB RAM recommended, 8 GB with the fast preset. No GPU needed | 8 GB RAM for the 4 billion model, 16 GB for the 12 billion, 32 GB for the 27 billion. Their own pages give three different minimums |
| Operating systems | macOS and Windows | Their own site conflicts for this SKU: the spec table says macOS and Windows, an article says Mac, Windows and Linux | Windows, macOS and Linux | Windows, macOS and Linux | macOS and Windows. Linux not mentioned | Their own site conflicts: the product page says Windows, macOS and Linux, the FAQ says Windows and macOS | Windows and macOS 13+. The PRO Navigator is Windows only | Windows only |
| Subscription | None | None. One-time purchase | None | None | None | None, though update packs are sold separately at $15 | None. Prompt packs are $8 each | None. The $249 and $469 tiers include a monthly cloud allowance that resets each month |
What actually ships, by name
Three of the seven competitors do not publish the names of the models they ship. BunkerAI describes six models only as Fast, Reasoning, Logic, Math, Expert and Creative. PortableMind says “the Qwen 3.5 roster” with no count. DrAcher names thirteen models using only its own aliases. This page makes an issue of that, so here is our own list in full.
| Role | Model | Size |
|---|---|---|
| Expert, top tier | Qwen3.6-35B-A3B-Instruct | 20.61 GB |
| Primary | Qwen2.5-14B-Instruct | 8.37 GB |
| Compatibility, low-memory fallback | Meta-Llama-3.1-8B-Instruct | 4.58 GB |
| Vision, Scout | Qwen2.5-VL-3B-Instruct | 1.80 GB |
| Companion | Phi-4-mini-instruct | 2.32 GB |
All five are Q4_K_M quantised GGUF files, about 37.7 GB together, and the vision model ships with its projector file. Which one runs depends on the host machine: the 8B is the low-memory fallback, the 14B is the everyday default, and the 35B is for machines with the memory to hold it. You can add your own GGUF models alongside these, limited only by space on the drive.
Where Off-Grid AI is genuinely different
The geographic layer. Three of the seven competitors ship a map you can pan and search, and one goes considerably further: PortableMind’s $199 PRO Navigator does turn by turn routing across the US road network with a GPS receiver in the box. That is a real navigation product, and on road routing with live positioning it does something we do not.
What no competitor mentions is the rest of the geographic stack: topographic contours, terrain layers, trail data, elevation, waypoints or coordinate entry. Every map claim on PortableMind’s site is streets and addresses. And none of the seven describes an AI that can query the map at all. Ask Off-Grid AI for the nearest mapped water source and the location comes from map data, then the treatment guidance comes from the cited library. The geographic fact is read from the map rather than guessed by the model.
That distinction is the argument. One competitor, DrAcher, states its answers carry numbered citations; the rest either describe retrieval in general terms or do not raise it. Meanwhile four of the seven market the opposite of restraint as the feature, selling “uncensored”, “unrestricted” or “unfiltered” output, and one of those sells it as a $40 jailbreak upgrade.
Off-Grid AI’s cited path works the other way round: it answers from retrieved evidence, names the document, and says so when the evidence is not there. Stopping the comparison there would be the same selective reporting this page criticises, though, because we ship an open chat mode as well. It is described in full below.
We have an unrestricted mode too, and here is what it actually does
Off-Grid AI has two ways to ask a question. Intel is the cited path: it retrieves first, names its sources, and declines when the evidence is not there. Companion is open chat with the loaded models, or any GGUF model you add yourself. Companion answers are not cited, and that is the point of it.
Companion also carries an Unrestricted toggle, so it is fair to ask how that differs from the marketing quoted above. It differs by exactly this much. Turning it on deletes the roughly seven lines of instruction we normally place in front of your message: be honest about uncertainty, decline clearly harmful requests, do not fabricate citations, and send survival and medical questions to the Intel tab. Those lines are replaced with an empty string. That is the entire mechanism.
What it does not do is unlock the model. Llama, Qwen, Phi and Gemma have refusals trained into their weights, and nothing here touches those. Ask one of them something it was trained to decline and it will still decline. There is no output filter being switched off either, because there is no output filter on that path to begin with. Our own acknowledgement screen says as much: for a genuinely unrestricted experience, use a community “uncensored” model.
Three things do change. The model stops redirecting survival and medical questions to Intel, so you get an uncited chat answer instead of being sent to the tab where answers are checked against sources. The medical disclaimer is suppressed. And you can supply your own system prompt, up to 4,000 characters, in the slot ours came out of.
It is off by default, requires a one-time acknowledgement you have to actively dismiss, shows a persistent “Unrestricted, unverified, uncensored AI” badge while it is on, and the setting is re-checked on the server for every message rather than trusted from the browser. The acknowledgement also tells you chats are stored unencrypted on the device. Intel is unaffected by any of it, and we do not sell the toggle as a paid upgrade.
Where one of the others is the better buy
A comparison whose author wins every row is one you should discount. Six cases where we are not the right purchase:
- You want the whole of Wikipedia offline. Trove ships full English Wikipedia, about 90 GB of it. We do not, and no amount of curation substitutes for wanting the encyclopedia. Our reasoning for leaving it out is below.
- You want road navigation with hardware. PortableMind’s $199 PRO Navigator ships a GPS receiver and does turn by turn routing across the US road network. We do not ship a GPS receiver. If driving directions with live positioning is the job, that is the product built for it.
- You are counting documents. DrAcher publishes 18,003 against our roughly 1,100, so on a straight file count theirs is bigger. Those are different axes, bulk government sourcing against a smaller reviewed corpus, but the number is the number. We also publish a page count, 161,286, because pages describe a corpus better than files do. No competitor here publishes one, so there is nothing to compare it against.
- You want a medical tuned model specifically. OffGrid AI Toolkit ships MedGemma alongside its general models. Ours is a general model reading a curated medical corpus, which is a different design.
- Price is the deciding factor. At $149.95 to $759.95, Off-Grid AI is the most expensive here at both ends of its range. OffGrid Drive is $49.95 but ships no hardware at all, and BunkerAI is $70 with a 128 GB drive included.
- You want an AI with no limits on what it will say. Four of these products market exactly that, one of them as a paid upgrade. Off-Grid AI is the wrong purchase for it, deliberately.
Why we do not ship Wikipedia
Offline Wikipedia is the single thing a competitor has that we get asked about most, so it is worth saying plainly that leaving it out is a decision rather than an oversight.
Intel makes a specific promise: every answer names the document it came from, and that document is stable and vetted. Wikipedia cannot carry that promise. It is community edited and mutable by design, which is a strength in an encyclopedia and a problem in a citation. A citation is only worth something if the thing it points at is still the same thing tomorrow.
The larger issue is what mixing does. Fold Wikipedia into the cited library and every answer inherits the reliability of the weakest source available to it, because the retrieval layer does not know which shelf it pulled from. That does not raise Wikipedia to the standard of a vetted manual. It lowers every vetted manual to the standard of an open wiki. Once a corpus is described as cited, the least reliable thing in it sets what the label is worth.
So we left it out. If a browsable offline encyclopedia is what you want on the drive, Trove ships one and we do not, which is why that is the first entry in the list above. What we will not do is put it in the cited library and carry on calling the cited library cited.
The one thing we tested ourselves
Every other line on this page comes from a company’s own website. This one does not, so we are flagging it as a different kind of evidence. We bought a BunkerAI drive and could not get it to run, on a Mac and on two separate Windows PCs. That is one unit on three machines. It is not a verdict on every unit they ship, we did not contact their support, and they may already have a fix. We report it because we ran the test.
A note on how to read any comparison like this one
Claims in this category decay quickly, and several of these companies do not agree with themselves. Trove was pre-Kickstarter when this was written, with a founder price of $69 against a $99 list. Docket’s own site contradicts itself on Linux support for its Survival model. BunkerAI contradicts itself twice, listing $70 on the page whose FAQ says $89, and naming Linux support on the product page but not in the FAQ. OffGrid AI Toolkit gives three different memory minimums across its own pages. PortableMind’s accuracy benchmark is self administered, which the company states plainly and to its credit. Three of DrAcher’s specifications appear only inside product images rather than in body copy, and its route planning is shown in an illustrative panel rather than listed as a feature. None of those discrepancies is stated here as a failing. They are stated so you check the date, then check the source.
