A coding agent, and the model we are building to run it.
Alvi Cloud makes Buraaq, an AI coding agent that runs in your terminal on Windows, Linux, and macOS. Today it calls third-party models. BuraaqGPT is our own 1.3-billion-parameter code model, trained from scratch, so the agent can run on infrastructure we control.
$ buraaq update --check current 0.9.1 channel latest, released 2026-09-23 windows buraaq-0.9.1-win-x64.zip linux buraaq-0.9.1-linux-x64.tar.gz macos buraaq-0.9.1-osx-arm64.tar.gz, buraaq-0.9.1-osx-x64.tar.gz up to date $
Buraaq, the agent
Buraaq reads, edits, and runs code from a terminal session. It ships as a single download for each platform and updates itself from a signed release manifest, so the same build runs on a laptop, a workstation, or a server without an installer.
It has had regular releases since August 2026. The current version is 0.9.1.
A second product, the Buraaq assistant, runs entirely on your own machine behind a default-deny approval gate and reaches you over Telegram. Both are built to run on BuraaqGPT.
Install it from a terminal. The script reads the release manifest, verifies the SHA-256, and puts buraaq on your PATH.
irm https://buraaq.alvi.cloud/install.ps1 | iexcurl -fsSL https://buraaq.alvi.cloud/install.sh | bash| Platforms | Windows x64, Linux x64, macOS arm64 and x64 |
|---|---|
| Interface | terminal application |
| Distribution | self-updating, SHA-256 verified releases |
| Models today | third-party APIs |
| Models next | BuraaqGPT, self-hosted |
BuraaqGPT, the model
BuraaqGPT is a decoder-only transformer trained from scratch on code. Every part of the pipeline is ours: the data build, the trainer, the evaluation, and the instruction-tuning data. Nothing is fine-tuned from someone else's weights.
The training corpus is 20 billion tokens: 80% permissively licensed code in 15 languages from The Stack v3, filtered with the Stack-Edu quality classifiers, 10% mathematics from FineMath, and 10% general text from FineWeb-Edu. Every document is exactly deduplicated and screened against 12,522 benchmark signatures from HumanEval, MBPP, GSM8K, and MATH, so evaluation results mean what they say.
The pipeline has been proven end to end on a 300-million-parameter model. The 1.3B configuration has been calibrated on rented GPUs, with checkpoint, resume, and distributed training verified.
Pretraining of the 1.3B model started on 20 September 2026 on Google Cloud. Every milestone is scored on the same fixed held-out data: loss on unseen code fell from 2.55 at 0.3 billion training tokens to 1.87 at 0.5 billion.
| Parameters | 1.30 billion |
|---|---|
| Architecture | 22 layers, 2048 hidden, grouped-query attention |
| Context | 4,096 tokens |
| Tokenizer | Qwen2.5-Coder, 151,936 rows |
| Corpus | 20B tokens, 80 / 10 / 10 code, math, text |
| Instruction data | 1.64M conversations, 681M tokens, test-verified |
| Licence policy | permissive SPDX allowlist, per-file provenance |
| Training | running since 20 Sep 2026 |
Where the work stands
- Data pipeline and trainerBuilt, tested, and validated on real data with a 300M model.
- 20-billion-token corpusCompleted 21 September 2026, audited part by part as it was built.
- Pretraining the 1.3B modelRunning since 20 September 2026 on one Google Cloud L4 GPU: 0.65 billion of the first 18.7 billion tokens trained as of 25 September.
- Instruction tuningOn the test-verified conversation set, so the model answers rather than completes.
- Buraaq runs on BuraaqGPTThe agent switches to the self-hosted model.
What the model needs now
Pretraining is running, but on a single L4 GPU at about 135 million tokens a day, so the first 18.7 billion tokens would take more than four months. On an H100 we measured 13,158 tokens per second: roughly 380 H100-hours would finish them in about two weeks, plus a fraction of that for instruction tuning and evaluation.
We are a self-funded company in Doha. If you run a cloud program, have idle GPU capacity, or want to sponsor an open code model with a shipping product behind it, we will publish the results and credit the support. Write to waseem@alvi.cloud.
About Alvi Cloud
Alvi Cloud is a self-funded software company founded in 2026 in Doha, Qatar, by Waseem Alvi. We build developer tools and the models behind them: Buraaq, a terminal coding agent with regular releases since August 2026, and BuraaqGPT, the code model we are training to run it.
| Company | Alvi Cloud |
|---|---|
| Founded | 2026, Doha, Qatar |
| Founder | Waseem Alvi |
| Products | Buraaq (agent), BuraaqGPT (model) |
| Releases | buraaq.alvi.cloud |