Question 1

What is Kespora?

Accepted Answer

Kespora is a desktop IDE for transformer architecture research — "Figma for LLMs". You build model architectures by dragging labeled blocks onto a visual canvas, then pretrain, fine-tune, evaluate, quantize, and chat with your models from the same app, and export clean PyTorch code at the end.

Question 2

What stages of the LLM lifecycle does it cover?

Accepted Answer

All of them, end to end, inside one workspace: Architecture design → Pretraining → Posttraining → Evaluation → Quantization → Chat → Export.

Question 3

Is Kespora open source?

Accepted Answer

Yes — the source is in the public repository on GitHub. License is TBD until a LICENSE file is added to the repo.

Question 4

Can I use it offline?

Accepted Answer

Yes. Pretraining, posttraining, quantization, evaluation, and chat all run locally. The AI assistant is the only feature that needs network access, and only if you wire it to an external provider.

Question 5

What PyTorch code does Kespora generate?

Accepted Answer

A complete, ready-to-run PyTorch project: model.py (GeneratedDecoderModel), attention.py, ffn.py, moe.py, config.json, and a README.md summary. The output is plain PyTorch — you can take the files and run them anywhere.

Question 6

What can the AI assistant do?

Accepted Answer

It can tell you about anything and do things and actions for you in the app — and you stay in control: every change goes through a review step before it touches your workspace.

Question 7

Which model templates ship with the app?

Accepted Answer

GPT-2, Llama (RoPE + RMSNorm + GQA + SwiGLU), Qwen, GLM (MLA-friendly), and a MiniMax-style 8-expert top-2 MoE template. Each is a starting point you can edit, save as a new workspace, and export.

Question 8

What posttraining methods are supported?

Accepted Answer

Five canonical routes: Simple (Base → SFT), Standard (Base → SFT → DPO/KTO), Unified (Base → ORPO), full RLHF (Base → SFT → Reward Model → PPO), and Reasoning RL (Base → SFT → GRPO/RLOO + verifier).

Log in to download