Thinks
It works the problem out before it answers.
Competition mathsAIME26
XCurOS 2.2 Max96.3
Qwen3.6-35B-A3B92.7
Contest programmingLiveCodeBench v6
XCurOS 2.2 Max88.2
Qwen3.5-397B83.6
XCurOS 2.2Max
An AI model that thinks, sees and gets the job done.
Use it free in the Playground, or inside your coding tools with an API key.
Arabic and English
Sees images, documents and video
A whole book at once
8 of 256 experts write each word
First on 34 of the 56 public tests.
Sees
It reads pictures, charts and video.
Understanding videoVideoMME
XCurOS 2.2 Max89.1
Qwen3.6-35B-A3B82.5
Reading chartsCharXiv RQ
XCurOS 2.2 Max81.2
Qwen3.6-35B-A3B78.0
Gets the job done
It plans, uses your tools and finishes the work.
Fixing real softwareSWE-bench Pro
XCurOS 2.2 Max59.6
Claude Opus 4.653.4
Working in a terminalTerminal-Bench 2.1
XCurOS 2.2 Max68.0
Claude Opus 4.665.4
Three ways to use it.
For engineers
Architecture and specs
Short name: XCurOS 2.2 Max · Name in the API: xcuros-2.2-max
Model ID XCurOS-2.2-A4B-Max; architecture
designation XCurOS-2.2-35B-A4B, shared with the base tier — a multimodal
Mixture-of-Experts causal language model with an integrated vision encoder: ~35B total parameters,
~4B active per token, and a context window of up to 1,048,576 tokens. It is post-trained from
XCurOS 2.2 A4B, the tier before it, and both tiers share everything listed here.
- Model IDs
XCurOS-2.2-A4B·XCurOS-2.2-A4B-Max- Type
- Multimodal causal LM (MoE) + vision encoder
- Parameters
- ~35B total · ~4B active per token
- Layers / hidden
- 40 layers · 2048 hidden
- Experts
- 256 routed · top-8 per token · FFN 512 · one shared expert always on
- Attention
- Hybrid Gated DeltaNet + Gated Attention
- Vision
- Integrated ViT — images and video
- Context
- 262,144 native · 1,048,576 with RoPE scaling · 512K on the API · 200K in the Playground
- Mode
- Thinking — always on the API, a switch in the Playground
- API
- OpenAI-compatible · Anthropic Messages
Thinking mode
Reasoning is what the hard answers are worth, so the API reasons before every answer. In the Playground it is a switch — the Think button by the composer — because a greeting does not need a reasoning pass and a proof does.
Small KV cache
Only the Gated Attention layers cache; the DeltaNet layers carry a fixed recurrent state. ~20 KB per token at fp16.
Use cases
On-premise copilots, document and chart understanding, OCR extraction, agentic automation, long-context analysis.
Max compared with A4B
XCurOS 2.2 A4B MaxLatest · what we serve
XCurOS-2.2-A4B-Max
Trained further from A4B: it set itself tasks, built the tests that grade them, and learned from the results. That makes it our strongest model for real engineering work, in the terminal and across a whole code base.
- Agentic coding, tool use, long-horizon planning
- Ahead of A4B on 52 of the 56 published benchmarks
- The model the Playground runs and the model card reports
XCurOS 2.2 A4BBase tier
XCurOS-2.2-A4B
The 2.2 model itself, with general training: reasoning, pictures, documents, tool calls and the full context window. Max is trained from it.
- Assistants, document and chart understanding, OCR, multilingual work
- Identical footprint and hardware bill to Max
- Measured on the same 56 benchmarks, under the same protocol
Both tiers are measured under the same protocol, so the two columns compare directly. Across all 56 published benchmarks Max leads on 52, is level on 3, and trails on one — MVBench, by 0.2. Eight of the widest gaps and the headline rows:
| Benchmark | XCurOS 2.2 A4B | XCurOS 2.2 A4B Max | Difference |
|---|---|---|---|
| Tool Decathlon | 31.8 | 49.1 | +17.3 |
| NL2Repo | 33.5 | 46.4 | +12.9 |
| Terminal-Bench 2.1 | 55.2 | 68.0 | +12.8 |
| OJBench | 46.2 | 54.3 | +8.1 |
| SWE-bench Pro | 53.2 | 59.6 | +6.4 |
| HLE | 20.7 | 24.7 | +4.0 |
| MMMU | 78.7 | 82.7 | +4.0 |
| SWE-bench Verified | 76.3 | 79.0 | +2.7 |
Where the parameters go
| Component | Total | Active — text | Active — image/video |
|---|---|---|---|
| Routed experts — 256 × 3 × 2048 × 512 × 40 layers | 32.21B | 8/256 → 1.01B | 1.01B |
| Shared expert — 3 × 2048 × 512 × 40 layers, always active | 0.13B | 0.13B | 0.13B |
| Attention — 10 Gated Attention + 30 Gated DeltaNet layers, incl. gates & conv projections | 1.28B | 1.28B | 1.28B |
| Embeddings (input + output, untied, vocab 248,320) | 1.02B | 1.02B | 1.02B |
| Vision encoder (depth 27, hidden 1152) | 0.44B | — | 0.44B |
| Total | ~35.1B ≈ 35B | ~3.4B | ~3.9B ≈ 4B |
The vision encoder runs only when an image or video is present, so a
text-only turn activates ~3.4B. The A4B designation refers to the multimodal figure.
Hardware
Memory is set by the total parameter count — every expert must be resident even though only 8 of 256 fire per token.
| Precision | Weights | KV @ 32K | KV @ 262K | Minimum GPU configuration |
|---|---|---|---|---|
BF16 | ~70 GB | ~0.6 GB | ~5 GB | 2× 80 GB (H100 / A100) |
Licensing
Every XCurOS Team product other than the open-weight releases is proprietary software — XCurOS 2.1 5B Instruct, XCurOS OCR, XCurOS 2.2 A4B, XCurOS 2.2 A4B Max and the serving stack that runs them. A valid paid licence is required for any use — personal, academic, evaluation or commercial. Access to the weights is gated on a licence reference.
XCurOS 2.2Max
An AI model that thinks, sees and gets the job done.
XCurOS 2.2 Max is an AI model: the part that understands what you ask and works out the answer. Chat apps, coding tools and company systems run on top of it — and it can run inside your own company, so your data never leaves.
Arabic and English
Sees images, documents and video
Can run inside your company
Works with the coding agents you already use
In plain words
The app is the screen. XCurOS is the mind behind it.
Every AI app has two parts: the screen you type into, and the model that understands you and writes the reply. XCurOS is the model: the part that does the thinking.
- 1You ask — in a chat, a coding tool or your own app.
- 2XCurOS reads it, thinks it through and writes the answer.
- 3The answer comes back to the app you asked in.
What people see
Apps people use
Chat
Coding tools
Company apps
Websites
This is what we built
The AI model
XCurOS 2.2 Max
Understands · Sees · Reasons · Writes
What it can do
One model. Many kinds of work.
Speaks your language
Reads and writes Arabic, English and many other languages.
Sees what you show it
Photos, screenshots, charts, scanned papers and video.
Reads long documents in one go
Up to 1M tokens (pieces of words) at once — several long books. The Playground runs at 200K, still a whole book.
Thinks before it answers
Works through maths, science and hard questions step by step.
Writes and fixes code
Builds features and fixes bugs inside the coding tools developers already use.
Uses tools to finish a task
Searches, runs commands and works with other apps until the whole job is done.
Measured, not claimed
Tested in public, against the best-known models.
960,000+
downloads on Hugging Face
Of our two open models, by the site's own count.
See the models ↗
Compare, one test at a time
XCurOS2.2 Max is the highlighted bar, with the A4B tier charted beside it · higher is better, and the highest is bold and underlined ·
-- means the score was not published
Full model card →
Private by design
Your data stays inside your company.
Many AI services send what you type to another company's computers. XCurOS can run on your own servers instead — even with no internet connection at all.
- Runs on your own servers
- Works with no internet connection
- Sends nothing back to us
Used worldwide
People open XCurOS from all over the world.
TOTAL VIEWS
0
pageviews — the home page, repeat visits included
TOKENS USED
0M+
tokens people have used on XCurOS so far — at least 409 million
UNIQUE
0
unique visitors — each person counted once, never twice, by Cloudflare
Get started
Three ways to start.
No sign-up
Try it
Ask XCurOS anything, right in your browser. Nothing to install.
Try it freeUse it in your coding tools
A key in a minute. Works with Claude Code, Codex and 8 more tools.
Get an API keyBring it to your company
Licences, pilots and a private setup on your own servers.
Talk to the teamXCurOS 2.2 A4B Max on 56 public tests
Scores on coding, maths, reasoning, pictures and languages, next to other well-known models. Higher is better. The full model card, as published with the model, follows.
The card is published in English.
How we measured
- Mode
- Thinking enabled
- Sampling
t=1.0 · top_p=0.95 · top_k=20- Precision
- BF16 · unquantized
- Measured on
- 2× A100 80 GB
Loading the model card…
We build AI that
companies can run themselves.
The XCurOS Team does AI research and engineering. We build the models and the software that runs them, so your data never leaves your company.
Five generations, six releases.
-
Founded
2022
XCurOS founded — secure AI systems for regulated environments.
-
Open weights8B Instruct
XCurOS 0.1
The first release. Weights published openly.
413,510 downloads → -
Open weights8B Vision-Language, BF16
XCurOS 1.2
Vision arrives: the model reads pictures.
548,352 downloads → -
Licensed5B Instruct
XCurOS 2.1
Tool calling, and the first deployments inside companies.
-
Licensed
XCurOS OCR
Document understanding — text, layout and tables off the page.
-
LicensedBase tier
XCurOS 2.2 A4B
A far larger model: video, and whole books at once.
-
LatestTrained further from A4B
XCurOS 2.2 A4B Max
The current model, trained for coding, tools and long tasks.
See the model →
Download figures are Hugging Face's own, read in August 2026.
What we hold to.
Security is the productIt runs inside your network, and your data stays yours.
Efficiency over spectacleOnly part of the model works at a time, so it needs less compute.
Claims you can checkWe publish every result, and how we measured it.
What we ship
Licences, pilots, partnerships.
Licences, evaluations, on-premise pilots and partnerships — the fastest route is a direct email.
Talk to the XCurOS team.
A question about your key or a tool, or a licence for your company? Write to us here, or pick another way below.
Before you write
Stuck connecting a tool?
Tell us the tool (for example Claude Code), your computer (Windows, macOS or Linux) and the error you see. Many answers are already on the API page.
Asking about a licence?
For a licence quote, it helps to include your deployment target (cloud, on-premise or air-gapped), expected concurrency, and whether you need vision and long context.
Evaluating the model?
Licensed users can reproduce every published figure with the bundled harness. Ask us for the evaluation notebook when you write.
Sending a security questionnaire?
Send them to [email protected] — the model runs entirely inside your perimeter, which answers most of them.
XCurOS service status
Is XCurOS working right now? Live status of the API, the Playground and our servers. This page runs on a separate system, so it keeps working when they do not.
Checking status…
Contacting the status service.
30-day uptime
—
XCurOS API
—30 days ago— uptimeToday
Playground
—30 days ago— uptimeToday
XCurOS Servers
—30 days ago— uptimeToday
Operational
Slow
Outage
Partly measured
Estimated
No data
—
—
—
—
Incident history
last 30 days Loading incident history…
Your account
One moment…
Use XCurOS in your own tools
Get your key, try it in code, or connect your tools in one line.
Get your key
Your key comes with your account. Create one at xcuros.com/signup: it takes a minute and it is free.
Your usage
Your own requests and tokens. Only you can see them.
Connect all your tools at once
Run this one line in your terminal, then paste your key in the page that opens. Every tool below is ready.
On a computer — Windows, macOS or Linux. Not on a phone.
curl -fsSL https://xcuros.com/vendor/i.sh | sh
How do I open Terminal?How do I open PowerShell?
Open Terminal: on a Mac press ⌘ Space, type Terminal and press Enter. On Linux press Ctrl Alt T.
Open PowerShell: press the Windows key, type PowerShell and press Enter.
Paste it in Terminal with ⌘ V and press Enter.
Paste it in PowerShell with Ctrl V and press Enter.
A page opens in your browser: paste your key and press Connect.
Or set up one tool by hand
Pick your tool to see its steps here.
Use it from your code
It works with the OpenAI and Anthropic SDKs: set the base URL, your key and the model below.
A Mixture of Experts model shares its work among many small experts. For each word, a router picks the few that fit best. So most of the model rests, and the answer still comes from everything it knows.
Base URL, key and model
- Base URL
https://xcuros.com/v1- Authentication
Authorization: Bearer YOUR_KEY- Model
xcuros-2.2
What your key includes
256KWith your email
xcuros-2.2 XCurOS 2.2, with a 256K-token context.
512KAdd your phone via Telegram, or pick a plan
xcuros-2.2-max XCurOS 2.2 A4B Max with a 512K-token context — and Claude's models in Claude Code, GPT's in Codex, at 1M.
10,000Limits
On Free, 10,000 requests and 200M tokens a month, with at most 2,500 requests and 50M tokens in any 7 days. The key works for 90 days; to renew it, contact the XCurOS team.
1MNeed more?
xcuros-2.2-ultra XCurOS 2.2 Ultra with a 1M-token context, on the Pro, Max and Custom plans. Paid plans start at $5 a month and go up to 100,000 requests and 2B tokens a month: see the plans.
API reference
Endpoints
| Endpoint | What it is for |
|---|---|
POST /v1/chat/completions | OpenAI Chat Completions: messages, streaming, tool calls, pictures (image_url) and JSON replies (response_format). |
POST /v1/responses | OpenAI Responses, as Codex and the newer OpenAI SDKs speak it. |
POST /v1/messages | Anthropic Messages, as Claude Code and the Anthropic SDKs speak it: streaming, tools and pictures. |
POST /v1/messages/count_tokens | How many tokens a request would use, before you send it. |
GET /v1/models | The models your key can use, with each one's context window. |
GET /v1/me | Your key's plan, its limits, what you have used and when each limit resets. Every answer also carries the same numbers in x-ratelimit-* headers. |
Models
| Model | Context | Notes |
|---|---|---|
xcuros-2.2 | 256K | XCurOS 2.2, what a key made with an email opens. Text, pictures and documents in; tool calls. |
xcuros-2.2-max | 512K | XCurOS 2.2 A4B Max. Opens with a verified phone number or any paid plan, with Claude's models in Claude Code and GPT's in Codex at a 1M-token context. |
xcuros-2.2-ultra | 1M | XCurOS 2.2 Ultra. Opens with the Pro, Max or Custom plan, beside Max and everything that comes with it. |
If a request fails
| Status | What it means |
|---|---|
400 | The request is longer than the model's context. Shorten it and send it again — the error gives both sizes. |
401 | The key is missing, wrong or expired. Check it above; if it has expired, contact the XCurOS team to renew it. |
403 | That model is not on your key. xcuros-2.2-max, and Claude's and GPT's models with it at 1M, open when you verify your phone number or pick a plan. |
413 | The request is larger than 48 MB. Send fewer or smaller attachments. |
429 | Too many requests at once on this key, or its allowance is used up. Wait as long as Retry-After says; the error names the limit and when it frees up. |
503 | The model is busy right now or starting up. Wait as long as Retry-After says and send it again; every OpenAI and Anthropic SDK does this by itself. |
5xx | The service is not answering. Try again in a few seconds; the Uptime page shows its state. |
Questions people ask
Is it really free?
Yes. The Free plan costs nothing: 10,000 requests and 200M tokens a month, on a key that works for 90 days. For more, see the plans.
Do I need to know how to code?
No. Connect sets up your tools for you: copy one line, paste it, press Enter, then pick your tools in the page that opens.
Which tools work with XCurOS?
Claude Code, Codex, GitHub Copilot, Antigravity, Kilo Code, OpenCode, Qwen Code, Grok Build, Cline and Continue. Any app that speaks the OpenAI or Anthropic API works too.
Can I use it on my phone?
Yes, in the Playground. Connecting coding tools needs a computer: Windows, macOS or Linux.
I lost my key. What now?
Enter the same email above. You get the same key back, not a new one.
Why is there a phone step?
It is optional. Your key works with your email alone, on xcuros-2.2 with a 256K context. Verifying your phone through Telegram opens xcuros-2.2-max with 512K, and Claude's models in Claude Code and GPT's in Codex at a 1M-token context.
How do I undo Connect?
Run the same line again and press Disconnect. Every file goes back exactly as it was.
All plans
XCurOS plan
One moment…
Sign in
One model. Pick your size.
Start free with your email on XCurOS 2.2. A verified phone or any paid plan opens XCurOS 2.2 Max; a bigger plan gives you more room.
Free
For trying XCurOS and learning. A key in a minute, with no card.
$0 / month
Get a free key
Requests
2,500a week
10,000a month
Tokens
50Ma week
200Ma month
Includes:
- XCurOS 2.2 with a 256K context, and Max with a verified phone
- Sees images, documents and video
- Claude Code, Codex, GitHub Copilot and seven more
- The Playground, free
- The key works for 90 days
Lite
For real work, a little every day.
$5 / month
Upgrade to Lite
Requests
5,000a week
20,000a month
Tokens
125Ma week
500Ma month
Everything in Free, plus:
- XCurOS 2.2 Max with a 512K context
- Claude's models in Claude Code, GPT's in Codex, at 1M
- The key stays on while you subscribe
- Kept on your existing key: nothing to reconfigure
- Help from the team by email
- Change your plan in any month
- No contract: stop whenever you like
Pro
For people who code, write and research with XCurOS every day.
$10 / month
Upgrade to Pro
Requests2.5× Lite
12,500a week
50,000a month
Tokens2× Lite
250Ma week
1Ba month
Everything in Lite, plus:
- XCurOS 2.2 Ultra with a 1M context, twice the room
- Ahead in line when the model is busy
- Help from the team on WhatsApp
- Long coding sessions without stopping
- Whole projects and big files in one conversation
Max
The most roomFor long coding sessions and agents that run for hours.
$20 / month
Upgrade to Max
Requests5× Lite
25,000a week
100,000a month
Tokens4× Lite
500Ma week
2Ba month
Everything in Pro, plus:
- First in line when the model is busy
- First to get new models
- A direct line to the team
- Room for agents that run for hours
- Enough for several tools working together
Custom
For companies that need limits sized to their traffic, a licence, or the model inside their own network.
Let's talk
Contact us
Everything in Max, plus:
- Limits agreed for your workload
- Can run inside your company
- Your data never leaves your company
- On-premise deployment and licensing
- Keys for your whole team, on one invoice
- Security questionnaires and a named contact
- Pilots and evaluations before you decide
Every key has a weekly and a monthly limit. The week is the last 7 days and frees up day by day; the month starts again on the 1st.
Pay for Lite, Pro and Max with PayPal, or with InstaPay in Egypt. The plan goes on your existing key.
Questions about plans
Compare every feature
| What you get | Free | Lite | Pro | Max | Custom |
|---|---|---|---|---|---|
| How much you get | |||||
| Requests a week | 2,500 | 5,000 | 12,500 | 25,000 | Agreed |
| Requests a month | 10,000 | 20,000 | 50,000 | 100,000 | Agreed |
| Tokens a week | 50M | 125M | 250M | 500M | Agreed |
| Tokens a month | 200M | 500M | 1B | 2B | Agreed |
| The model | |||||
| XCurOS 2.2 Max with a 512K context | With your phone | Yes | Yes | Yes | Yes |
| Claude's models in Claude Code, GPT's in Codex, at 1M | With your phone | Yes | Yes | Yes | Yes |
| XCurOS 2.2 Ultra with a 1M context | No | No | Yes | Yes | Yes |
| First to get new models | No | No | No | Yes | Yes |
| Thinks before it answers | Yes | Yes | Yes | Yes | Yes |
| Sees images, documents and video | Yes | Yes | Yes | Yes | Yes |
| Arabic and English | Yes | Yes | Yes | Yes | Yes |
| Where you use it | |||||
| The Playground | Yes | Yes | Yes | Yes | Yes |
| Claude Code, Codex, GitHub Copilot and seven more | Yes | Yes | Yes | Yes | Yes |
| OpenAI and Anthropic APIs | Yes | Yes | Yes | Yes | Yes |
| When the model is busy | |||||
| Your place in line | Normal | Normal | Ahead | First | First |
| Your key | |||||
| How long the key works | 90 days | While you subscribe | While you subscribe | While you subscribe | Agreed |
| Check your usage at any time | Yes | Yes | Yes | Yes | Yes |
| Help and your company | |||||
| A live status page | Yes | Yes | Yes | Yes | Yes |
| Help from the team by email | No | Yes | Yes | Yes | Yes |
| Help from the team on WhatsApp | No | No | Yes | Yes | Yes |
| A direct line to the team | No | No | No | Yes | Yes |
| Can run inside your company | No | No | No | No | Yes |
| A licence for the model | No | No | No | No | Yes |
| Keys for your whole team, on one invoice | No | No | No | No | Yes |
| Security questionnaires | No | No | No | No | Yes |
Which plan should I pick?
Start with Free. When your work reaches its limits, move to Lite. Pro gives you 2.5 times Lite's requests and Max five times. Your key stays the same when you move.
How do I pay?
Choose your plan and pay with PayPal, or with InstaPay in Egypt. The plan goes on your existing key, so there is nothing to reconfigure. If you need another way to pay, write to us from the contact page.
Do I need a new key to change plans?
No. Your key stays the same and its limits change, so every tool you have connected keeps working.
What is a token?
A piece of a word. Both what you send and what the model writes back are counted.
What is a context?
How much the model can hold in mind at once in one conversation. A bigger context means longer conversations and bigger files before it starts to forget.
What happens when I reach a limit?
The API answers 429 and says which limit was reached and when it frees up. Nothing is cut silently and no other model answers in its place.
How the limits work
Requests a month Every request, whatever its size. The count starts again on the 1st of each month, UTC.
Tokens a month Input and output tokens in the calendar month. The count starts again on the 1st of each month, UTC.
The weekly limit Requests and tokens over the last 7 days. They free up day by day as older use leaves the window.
Requests at once One key may run up to half of the model's capacity at the same moment, so nobody can take all of it. Past that the API answers 429 with Retry-After.
When the model is busy A request waits a few seconds for a free place. If none frees up, the API answers 503 with Retry-After — never a different model in our place. Every OpenAI and Anthropic SDK retries this by itself.