SULTAAN.AI

Roadmap

Five phases, climbed in order

The roadmap in the repository is the ground truth for what is in scope; the longer vision document is a north star, not an instruction manual. Phases are counted in working sessions, not dates, so no release date is promised here.

The stair

Where each phase stands

  1. Foundation Complete

    • Repository skeleton, the Rust Backend trait and the Python package.
    • Hello-world text on both backends.
    • Perplexity baseline, pre-commit hooks and the first five ADRs.
  2. Fine-tune, evaluate, act Mostly complete

    • Agent loop with a Python tool: done.
    • Evaluation suite (perplexity, MMLU, TruthfulQA, hallucination): done.
    • QLoRA pipeline: scaffolded. The first trained adapter for a 7B open model is still to come.
  3. Memory, retrieval, sandbox Code complete

    • Persistent memory, hybrid retrieval, SOUL.md, knowledge graph, grounding verifier and citations, all unit-tested.
    • Live demonstrations of memory and the sandbox wait for a suitable host, and the sandbox is not yet wired into the command line.
  4. Pretraining Current

    • From-scratch trainer built; a CPU smoke run took loss from 5.49 to 4.33.
    • First-contact run on a 6 GB laptop GPU in July 2026.
    • Next: a roughly 200M-parameter run on a rented GPU, aiming for perplexity under 30 on English text.
  5. Product and interface Planned

    • A desktop app, voice, and a distilled smaller model.
    • Signed multi-architecture images, REST and gRPC APIs, model signing.
    • Edge guides for Raspberry Pi and Jetson, and a custom SULTAAN.AI licence.

The ladder

Model sizes, rung by rung

What each tier needs to train from scratch. Training is limited by GPU memory, not system RAM.

Training ladder
TierApprox. sizeGPU memory to trainWhere it runs
T0~20M parameters~1 GBAny laptop; CPU smoke test done
First contact~34M parameters4.4 GB measured6 GB laptop GPU; run done July 2026
T1~200M parameters~5–6 GBBorderline on a laptop; one 12–24 GB GPU
T2~460M parameters~12 GBOne 24 GB GPU
T3~1.5B parameters~30–40 GBOne 40–80 GB data-centre GPU
T4+3B and up48 GB or moreA multi-GPU cluster

Sizes are approximate. They assume the 152K-token Qwen2.5 vocabulary, whose embedding table dominates small models; a smaller custom vocabulary shrinks them. Training a 30B model from scratch is a cluster-scale job, never a single rented machine.

Boundaries

Never in scope, unless the roadmap changes

The project builds on proven foundations instead of reinventing them:

  • No home-made deep-learning framework: PyTorch and Candle.
  • No home-made tokeniser library: Hugging Face tokenizers.
  • No home-made vector database or database: an existing vector store, SQLite and Postgres.
  • No from-scratch FlashAttention: the engines already provide it.

Changing a phase boundary takes a written ADR with a blast-radius section, not a quiet refactor.

Read the first-contact storyContact via askshiva.in