Blog

Notes from the climb.

Notes from KTwo AI Labs on self-training and multimodal models, LLM API pricing, AI for Hungarian law, and how LE PEAK, our free multi-model chat, is built.

Follow along with the RSS feed.

  1. Research

    Mixture of experts, explained: why 120B can run like 5B

    How mixture-of-experts AI models work, what names like 30B-A3B and 550B-A55B mean, and the trade-off between total and active parameters in memory, speed and cost.
  2. Útmutató

    Mi az a token? Tokenizálás, és miért drágább magyarul

    Az AI-modellek tokenekben olvasnak, nem szavakban. Hogyan működik a tokenizálás, és miért kétszer annyi token ugyanaz a szöveg magyarul, mint angolul.
  3. Guide

    What is a token? Tokenization, and why Hungarian costs more

    AI models read tokens, not words. How tokenization works, how to estimate token counts, and why the same text in Hungarian takes about twice the tokens of English.
  4. Guide

    What is a context window? AI memory limits, explained

    A model’s context window is how much text it can read at once. What it counts, why long contexts cost memory and money, and why bigger isn’t always better.
  5. LE PEAK

    Attach PDFs, images, video and audio to a free AI chat

    LE PEAK now reads your files: ask free AI models about PDFs, photos, screenshots, videos, code and audio. Which models read what, the limits, and what stays private.
  6. LE PEAK

    LE PEAK: a free AI chat with 23 models in one thread

    Chat with DeepSeek, Gemma, gpt-oss, Nemotron and more for free, twelve models without an account. How LE PEAK works, its daily quotas, attachments and what it keeps.
  7. Research

    What is a self-training AI model?

    Most language models stop learning the day they ship. What self-training means, how it differs from fine-tuning and RAG, and why it is hard to get right.
  8. Pricing

    LLM API pricing explained: cost per million tokens

    How per-million-token pricing works, why output costs more than input, worked monthly cost examples, and the hidden multipliers that inflate an LLM bill.
  9. Research

    Natively multimodal models, explained

    Some AI models see and read with one network trained on everything from the start; others have vision bolted on. How the two designs differ and why it matters.
  10. Research

    Why 9B? The case for compact AI models

    What a model’s parameter count means for memory, speed and cost, how to read names like 30B-A3B, and when a compact 9B model is the right choice.
  11. Hungarian law

    AI for Hungarian law: introducing KTwo Mantle 0.5

    Why Hungarian law needs its own AI model, what makes legal Hungarian hard for general models, how to judge a legal model, and where KTwo Mantle 0.5 stands.
  12. Magyar jog

    Mesterséges intelligencia a magyar jogban: a KTwo Mantle 0.5

    Miért kell a magyar jognak saját AI-modell, mi nehéz a jogi nyelvben az általános modelleknek, hogyan értékeljünk jogi modellt, és hol tart a KTwo Mantle 0.5.
  13. Engineering

    Per user, per model, per UTC day: metering a free AI chat

    How LE PEAK enforces daily AI quotas with one atomic SQL statement, reserves room for streamed replies, and rate-limits guests without storing IP addresses.
  14. Engineering

    Drawing K2 live in WebGPU

    How ktwo.dev draws a night climb of K2 with WGSL shaders: a scroll-driven camera, a contour map, a footer engraving, and fallbacks for browsers without WebGPU.
  15. Guide

    AI model glossary: 37 terms in plain words

    Plain-language definitions of the terms on AI model spec sheets: parameters, tokens, context windows, mixture of experts, quantization, multimodal, self-training.