KGP Talkie

Qwen shipped three 27B models in about six months: Qwen 3.5 27B in February, Qwen 3.6 27B in April, and Qwen 3.8 27B in August. Each release came with bigger benchmark numbers. A simple question is left unanswered by all of them: did the architecture chan

Every local model comparison runs each question once, writes the score in a table, and calls it a benchmark. One run. That number hides the thing we care about most: does the model give us the same answer the second time we ask? consistency@10 is the metr

Speculative decoding lets a small helper guess the next few tokens, and the full model checks all of them in one pass. In simple words, the model drafts ahead, and it keeps only the guesses it would have written anyway. Qwen 3.8 27B ships that helper insi

Every tokenizer has one dial you can turn: how many tokens to keep in the vocabulary. Turn it up and each sentence needs fewer tokens. Turn it down and the model gets lighter. You cannot have both. This is the central trade-off of tokenization, and it qui

Easy questions cannot separate local models any more. Ask a modern 27B model to write a palindrome checker and all of them get it right, so the score tells us nothing. In this blog, we will push three of them until they break. We built sixteen hard proble

Two frontier models shipped inside 24 hours of each other. xAI released Grok 4.6 on August 12, 2026, aimed at long-running agents and interactive work. Google released Gemini 3.7 Flash on August 13, 2026, calling it their "most intelligent workhorse model

Two new open models came out this month, and both fit under 35B parameters. NVIDIA released Nemotron 3.5 Lightning on August 11, and Meta released Muse Glimmer 30B a little earlier. Everyone compares these models using vendor benchmark tables. I wanted to

Open any tokenizer's vocabulary and you will find strange entries near the top: CLS, SEP, PAD, UNK. They are not words. Nobody writes them. Yet without them, a Transformer cannot do its job. In simple words, these are control tokens. They carry instructio

We have now met all three subword methods. BPE glues the most frequent pair. WordPiece glues the pair that best predicts the text. SentencePiece skips word splitting and reads the raw stream. Three methods, one job. So how do you tell them apart in the wi

BPE and WordPiece both quietly assume one thing: that you can split text into words on spaces first. But many languages do not use spaces at all. SentencePiece, the method behind T5 and Llama, throws out that assumption and treats text as one raw stream o