Coaade Mind 1.0
The decisive personal intelligence that lives on your device.
Not a tourist in the cloud. A resident expert on your phone, laptop, and desktop, that calls your tools, makes the call, and keeps your data where it belongs: with you.
Frontier models are brilliant tourists in the cloud that hedge when you need an answer. Coaade Mind 1.0 is the decisive, straight-talking resident expert that lives on your device, is best-in-class at decisions and comparisons, and becomes your AI company brain across Notion, Gmail, and the rest of your data - so a leader can actually decide.
Coaade Mind 1.0 is in planning (target release Q2 2027). The capabilities, benchmark targets, and hardware figures below are the goals and expected estimates we are building toward.
What it is
Coaade Mind 1.0 is a new kind of AI from Coaade Inc: a Personal Intelligence, and a first step toward basic Physical AI. It is a distinct product line from the Coaade Code engineering models. Where Coaade Code is a specialist software engineer, Coaade Mind is the always-available assistant that lives with you across every device you own.
It is deliberately small - a 1.5B-parameter model, deeply over-trained and hyper-specialized so it beats far larger cloud models on the tasks that actually matter to a personal AI. It carries a 1,000,000-token context, so it can hold an entire project, inbox, or company knowledge base in mind at once. And it runs fully on-device, on Apple, Windows, and Ubuntu, from phone to desktop, powered by the Coaade StrataFlow engine.
It is built on the GTV (Generate, Test, Verify) framework created by Ammar Tahir, CEO of Coaade, so its answers are grounded, checked, and honest rather than confidently wrong.
1.5B
parameters, deeply over-trained
1,000,000
token context window
100% on-device
phone to desktop
What it is best at
Coaade Mind is not a jack-of-all-trades that is master of none. It is purpose-built to be exceptional at a focused set of things that make a personal AI genuinely useful.
Expert tool calling
A best-in-class tool caller: native function calling plus full MCP-client support, parallel and sequential calls, agentic think-act-observe loops, smart tool selection, and graceful error recovery. It does not just suggest what to do - it does it.
Decisive decision-making
Most AI hedges. Coaade Mind decides. Its engine frames the problem, lays out options, sets criteria, scores them, recommends, and stress-tests with sensitivity analysis - then commits with the reasoning and tradeoffs on the table. No endless 'it depends'.
Your AI company brain
Connect Notion, Gmail, Calendar, Slack, Drive, your CRM, and your databases. Coaade Mind pulls them into its million-token context and turns scattered data into cross-source insights, summaries, briefings, and grounded recommendations.
Visual and interactive answers
Where a widget beats a wall of text, Coaade Mind generates interactive, real-time UI inline: live charts, sortable tables, decision trees, board widgets, forms, maps, and dashboards - a safe declarative spec your device renders into something you can click and act on.
It learns you
Serving exactly one user on that user's own device, Coaade Mind personalizes over time - learning your habits, your voice, and your corrections from your feedback - and never ships any of it off your device.
Deep domain knowledge
Core expertise in personal intelligence, finance, marketing, and copywriting, on top of a broad base spanning translation, IT and computer science, ML and AI, physics, psychology, literature, film, music, geography, food, startups, math reasoning, and far more.
How good it will be (targets)
Coaade Mind 1.0 is tiny next to frontier cloud models, yet it is designed to beat them where it counts. The figures below are the aspirational targets we are engineering toward for the shipped 1.5B flagship. They are goals, not measured results.
| What we measure | Coaade Mind 1.0 target | Frontier cloud generalists |
|---|---|---|
| Berkeley Function-Calling (BFCL v3) | 92-95% | ~85-88% |
| Parallel tool calls (exact match) | 90%+ | 74-80% |
| Agentic tool loop (task success) | 85%+ | 70-78% |
| Decision quality (Coaade Decide-Bench) | 88%+ | 72-75% |
| Option-comparison faithfulness | 90%+ | 73-78% |
| Personal-intelligence eval | 90%+ | 68-72% |
| 1M-context needle-in-haystack | 99%+ | not supported at 1M |
| Finance reasoning | 84%+ | 80-82% |
| Runs fully on-device, phone to desktop | Yes | No |
The thesis
A tiny, deeply over-trained model (1.5B parameters, 400 tokens per parameter) that is hyper-specialized for tool calling, decisions, and personal intelligence - and that runs where frontier models simply cannot (your own phone and laptop) - wins the tasks that matter to a personal AI even while being far smaller.
How much RAM and GPU you will need
Here is the part people always ask about: you do not need a datacenter. Coaade Mind runs on hardware you already own. A small GPU or Apple Silicon makes it snappy, but it also runs CPU-only on a modern laptop. The figures below are expected estimates.
| Your device | Example | Active RAM (StrataFlow) | Expected speed | Context you can use |
|---|---|---|---|---|
| Phone | iPhone / flagship Android | 2-3 GB active | ~30-45 tok/s | up to 128K (windowed) |
| Tablet | iPad Pro | 2-3 GB active | ~45-60 tok/s | up to 256K |
| Thin laptop | MacBook Air, ultrabook | 2-4 GB active | ~60-80 tok/s | up to 512K |
| Pro laptop | MacBook Pro, RTX laptop | 3-5 GB active | ~80-110 tok/s | full 1,000,000 |
| Desktop GPU | RTX 4090 / 5090 | 3-6 GB active | ~120-180 tok/s | full 1,000,000 |
| Workstation | Mac Studio / dual-GPU | 4-8 GB active | ~150-220 tok/s | full 1M + multi-session |
StrataFlow is what makes these numbers possible: it keeps only the hot experts and trunk resident and streams the rest on demand, so the active memory footprint stays small - just a few GB, even on a phone - while the hot path runs fast. Low RAM used, high speed delivered: no datacenter, no GPU rental. A dedicated GPU pushes the numbers higher still.
How small is the download?
The shipped 1.5B flagship is about 1.0 GB on disk at Q2, holding a 2-3 GB working set for everyday use (6-8 GB when you fill the full 1M-token context). The optional larger variant (a ~3B-total Mixture-of-Experts build) is about 1.9 GB on SSD and still keeps only a 2-3 GB active working set, because StrataFlow streams the cold parts from your SSD on demand - how even a 16 GB laptop reaches the full million-token context.
Bottom line
If you have a reasonably modern phone, tablet, or laptop, you can run Coaade Mind. No cloud GPU rental, no datacenter, no per-token bill.
How it is built (for the curious)
This is the one-time training cost, not what you need to run it. Coaade Mind 1.0 is pre-trained on a reserved block of 8x NVIDIA H200 GPUs (AWS p5e.48xlarge) inside a committed $150,000 pre-training envelope, then post-trained (supervised fine-tuning and reinforcement learning) on 8x H100-class nodes. The total program budget for the staged 1.5B plan is $1,590,000. You pay none of that to run the model: once it is on your device, it is yours.
How it is deployed
Coaade Mind runs natively across the platforms you use.
Apple (macOS, iOS, iPadOS)
Metal / MLX with unified memory - the best-case StrataFlow target
Windows
CUDA for NVIDIA GPUs, with a DirectML fallback
Ubuntu / Linux
CUDA, ROCm for AMD, or a CPU path for headless servers
Under the hood, the StrataFlow engine streams the model's experts and context across VRAM, RAM, and SSD, keeping the hot parts in the fastest memory and streaming the rest on demand. The result: large-model behavior from a small active footprint, on the hardware you already own.
Your data stays yours
Coaade Mind is on-device by default. Your personal and company data stays on your device - not pooled, not shipped off to a datacenter without your explicit, per-category consent.
Private by design
Your data does not leave your hardware.
Free after download
No per-token cloud fees, ever.
Works offline
On a plane, in a tunnel, anywhere - no network round-trip.
A cloud model structurally cannot make that promise. Coaade Mind can, because the whole model lives with you.
Get Coaade Mind 1.0
Coming to Apple, Windows, and Ubuntu, from your phone to your desktop. A decisive, private, expert intelligence that is genuinely yours. Register your interest and be first in line when it ships.