News & guides
Unicode and Tokenization: Where Text Handling Goes Wrong
Emoji that cost five tokens, accented text that costs double, and truncation that produces broken characters. How Unicode meets the tokenizer, and what breaks.
· 9 min read
Vocabulary Size Trade-offs in Language Model Tokenizers
A bigger vocabulary shortens sequences and costs parameters and rare-token quality. What the dial actually controls and how it reaches your bill.
· 9 min read
Where Quantised Models Actually Lose Quality
Quantisation degrades capabilities unevenly rather than uniformly. Which ones erode first, why a benchmark delta hides it, and how to measure the loss on your own work.
· 9 min read
Context Windows: Why More Is Not Always Better
A million-token context window sounds like the end of chunking. In practice models get slower, pricier and less accurate long before you fill it. Here is why.
· 8 min readEmbeddings Explained: Vectors, Similarity and Where It Breaks
Embeddings turn text into vectors so you can search by meaning instead of keywords. Here is how they work, what similarity really measures, and the failure modes to expect.
· 9 min read
Fine-Tuning vs Prompting vs RAG: How to Adapt a Model
Three ways to make a general model do your specific job, with very different costs. Here is a decision rule based on what each one can and cannot actually change.
· 9 min readHow LLMs Actually Work: Next-Token Prediction, End to End
A large language model does one thing repeatedly. Here is what happens between your prompt and the first token out, and what that mechanism explains about behaviour.
· 9 min read
LLM Latency: Why Time to First Token Is the Number to Watch
LLM latency is two different problems wearing one name. Understanding prefill and decode tells you which knob to turn when a request feels slow.
· 9 min readMixture of Experts: Why Trillion-Parameter Models Are Fast
MoE models decouple parameter count from compute per token, which is why a trillion-parameter model can be cheap to serve. Here is the mechanism and what it costs you.
· 8 min readModel Distillation Explained
How a small model inherits the behaviour of a much larger one, what gets lost along the way, and why distillation is the reason cheap models got good so fast.
· 8 min readMultimodal Models Explained: How Images Become Tokens
A multimodal model does not see an image. It sees patch embeddings projected into the same space as text tokens. What that implies for cost and accuracy.
· 8 min readOpen Weights vs Open Source: The Distinction That Matters
Downloadable weights are not the same thing as open source. Here is what the licences actually permit, and which differences change what you can ship.
· 9 min readReady to try it?
Every model behind one key. Start free and upgrade when it sticks.