Our vision · Make a safe AI world

Fast, private AI on the machine you already own

We believe the best place to run AI is on your own computer: close to your files, under your control, and on hardware you have already paid for. AlphaLlama exists to make that fast enough that you never need to send your work to someone else's datacenter.

Run locally. Save energy by running faster.

Generating text on a GPU is limited by how fast it can read the model's weights from memory, so the card draws much the same power whether the software is fast or slow. The energy an answer costs is that power multiplied by the time it takes. Finishing the same answer in less time uses less energy, on a machine that is already on your desk, with no network round trip and no datacenter cooling.

That is why we measure speed so carefully. On InferArena, AlphaLlama runs Qwen3.8-27B at 659 tokens/s on a single RTX 3090 and the ~125B Qwen3.8-Flash-Next at 602 tokens/s on a single RTX 5090 (fast mode, best task). Expert offload lets a model larger than the GPU run from system memory, so a 16 or 24 GB card can do work that used to need a server.

Privacy and safety by design

Your data stays with you

Prompts, documents and answers are processed on your own machine. Nothing has to be uploaded to use the engine, and it keeps working without an internet connection once the model is downloaded.

No account, no license key

AlphaLlama is free to download and run. There is nothing to sign up for and no key that ties your use to us.

Open source you can inspect

The engine is open source under Apache 2.0. Anyone can read the code, build it themselves and check what it does with their data.

Honest output and honest numbers

Exact mode produces the same text as standard greedy decoding; faster modes that can change the text are always labelled. Our benchmark publishes its rules, failures and per-prompt results, and invites corrections.

Our history

  1. 6 May 2026

    AlphaLlama Drop 1

    • •Install and run an open model on your own machine in about 30 seconds.
    • •Long-context retrieval over very large document collections on a single RTX 3090.
    • •Windows, macOS and Linux. Your data stays on your device.

  2. 2 November 2026

    AlphaLlama v2 (2.0.0)

    • •The AlphaInfer engine backend: one captured GPU graph per decode round, speculative decoding with MTP and n-gram copy drafts, and expert offload for mixture-of-experts models larger than the GPU.
    • •Exact mode, with output identical to standard greedy decoding, and a labelled fast mode.
    • •Qwen3.8-Flash-Next, a ~125B model, at 602 tokens/s on one RTX 5090 (fast mode, best task).
    • •Open source under Apache 2.0, measured in the open on InferArena.

  3. 3 November 2026

    AlphaLlama v2 (2.0.1)

    • •Fixes a server crash on certain long Qwen3.8-Flash-Next prompts.
    • •The API reports the model file's name as the model id; packages for Linux, Windows and macOS.

Contact us

Questions, ideas, corrections to our benchmarks, or interest in working with us: write to developer@alphachat.com