A real chat model, on a stack you can verify from the first byte.

This demo runs a real, modern instruction-tuned chat model — SmolLM2-360M-Instruct (2024) — on Helix, a software stack that rebuilds itself from 299 hand-typed bytes, so every layer of it can be audited instead of trusted. It also runs the classic GPT-2-XL and more — switch any time. Every output is checked, token by token, against an independent referee. Pick your door:

THREE DOORS

Choose your depth

FOR EVERYONE

The guided run →

Watch the model think, step by step, with every stage explained in plain language. Click anything to learn more. Start here.

journey.html FOR TINKERERS

The expert playground →

The full instrument panel: branching chats, the raw op-stream, kernel sources, presets, deep-think modes, exports.

index.html FOR SKEPTICS

The proof →

The four recorded checks behind the claim — rebuild, referee, repeatability, attestation — each one click from its evidence.

dashboard.html

HONESTY UP FRONT

What this demo is — and isn't

IT IS

· The real, unchanged public SmolLM2 (and GPT-2), running on kernels compiled from Helix source by a compiler that rebuilds from 299 bytes powered by Helix

· Verified: every model matched an independent oracle token-for-token (8/8 for the SmolLM2 chat, 25/25 for GPT-2-XL), gated fail-closed

· Honest about its sources: the badge always says LIVE, REPLAY (a real recorded run) or PREVIEW (mock)

IT ISN'T

· A frontier assistant. The default SmolLM2-360M-Instruct is a small 2024 chat model — it answers simple prompts, but at 360M it's tiny, so keep expectations modest. (The classic GPT-2-XL here is a 2019 base completion model that only continues text.)

· A speed demo. The small SmolLM2 chat answers in a few seconds; the larger GPT-2-XL runs ≈10 s/token by design — the pitch is trust, not raw speed (it gets faster as Helix develops).

· Verified below PTX. One closed NVIDIA step is trusted-once — disclosed, never hidden.

DON'T TAKE OUR WORD

Re-check the core yourself — one command, one minute

# clean checkout · CPU-only · no GPU, no weights · fail-closed git clone https://github.com/Questeria/helix && cd helix bash scripts/reproduce_trust.sh # rebuilds the entire ladder from the 299-byte root and asserts: seed 9837db12 · fixpoint 0992dddd · gcc-DDC K1 84363adb

The model legs additionally need the public HuggingFace weights and an independent oracle — the repo's runbook walks through both tiers honestly.

Honest residuals: fp32 · verified to PTX, not SASS · single GPU (sm_86) · base models, not assistants · the oracle shares the model's spec. Every number on this page is a recorded, committed result. start here · guided run · expert · proof · models