Why a model drops the middle of a long document
On 500 pages of docs, full context scored about 34%. A small model extracts facts before the expensive one answers.
Stuzhuk Lab — Chemistry of Code
Chemistry of Code
← All posts
Tag
llm
All blog posts with this tag.
Neural network quality metrics: how to know the model actually works
How to choose neural network quality metrics: classification, computer vision, OCR, LLM and RAG, calibration and production indicators — what to measure so the model truly works in the product.
Cohere quietly shipped a 218B MoE model just for translation
North Small Translate: 218B parameters, 25B active, 50 languages, WMT26 83.60. Open weights under CC BY-NC, production via Model Vault — specialists over chatbots.
Qwen 3.8 27B on Cerebras at ~1500 tokens/s
Qwen 3.8 27B on Cerebras public endpoints: ~1500 tokens/s, 64k/128k context, original unpruned weights in production.
Private Mistral codebook generator: a plan, not a recipe
Design a local qualitative-research workflow with Mistral Small 3.1, Ollama, and FastAPI—without overstating what AI can prove.
CMP 90HX for LLM: why a mining GPU slows down and what “six-seven” fixes
MoE tolerates CMP 90HX; dense models don't — compute paths are capped. Driver mod via V67 nearly doubles throughput but doesn't erase mining-card limits.
How LLMs Find Knowledge: From Tokens and Transformers to RAG
How text becomes tokens and vectors, what a Transformer actually does, why training is not RAG, and how to grow a Markdown knowledge base into retrieval.
How to evaluate LLMs before production: GitHub secret scanning lessons
Start with the product decision and guardrails, then the model—offline eval as integration tests plus error analysis.
How LLMs Are Trained: Pretraining, Fine-Tuning, LoRA, RAG, and Agents
A practical map of LLM adaptation: what changes model weights, what plugs in knowledge and tools, and how to choose an approach without “fine-tuning on PDFs.”
Ukrainian LLM Leaderboard: ranking models on Ukrainian
Ukraine’s Ministry of Digital Transformation, UCU, and lang-uk open an Hugging Face leaderboard for LLMs on Ukrainian: translation, summarization, retrieval, ZNO.
Cheap LLMs for agents: why token savings raise task cost
Why distilled models spawn extra tool calls, how benchmarks mislead, and which metrics beat price-per-million for agent TCO.
NanoGPT Speedrun: 153 agent runs across frontier models
Prime Intellect’s nanoGPT speedrun: 18 models, validated top score for Kimi K3kimi-code, plus 41 full agent trajectories.
Self-hosted LLM for 25 developers: prefix cache beats raw model choice
vLLM on Blackwell: 96.9% prefix cache hit rate, 452:1 input/output, and why teams keep inference in-perimeter for control—not token savings.
Language Model Quality Testing: What the Tests Are and What They Actually Measure
LLM quality checks: public task sets, arenas, model judges, safety, product regression, and online signals — what each test measures and what it cannot see.
Semantic context for LLMs: how LSP helps a model understand code
Part 4 of a completion series: Language Server Protocol supplies types, definitions, and references — a layer on top of repository text search.
Claude system prompts: lessons from Haiku 3 to Opus 5
Anthropic publishes system prompts: not magic personas, but decision rules, tools, uncertainty handling, and UX for long agent tasks.
LLMs overturn old math claims—verification is the bottleneck
Counterexamples and proofs from models, a sandbox escape, Lean certificates: don't treat chain-of-thought as faithful reasoning.
LLM Gateway & FinOps in 2026: Model Routing, Token Budgets, and Cost Control Without Sacrificing Quality
A practical 2026 guide to LLM gateways and FinOps: request taxonomy, model routing, token budgets, caching caveats, failover, multi-provider contracts, chargeback, security, observability, and a 90-day rollout plan.
OpenAI brings frontier AI to US scientific research
OpenAI’s National Science Initiative funds model access for labs and universities, but validation, governance, and expert review remain decisive.
Kimi K3: What Open Weights Mean for Developers
Moonshot AI has released Kimi K3 weights. Here is what its MoE design, 1.4 TB footprint, and licence mean before adoption.
How to Select an Enterprise LLM: Criteria Over Rankings
How to select a language model for enterprise AI using scenarios, security, cost, and infrastructure requirements.
Open LLMs in production: 8 lessons from llama.cpp, Gemma, and Qwen
Eight practical lessons for running open LLMs: inference control, task-specific model selection, prompt templates, and structured output.
A 744B AI Model in 25GB of RAM: Inside Colibri
Colibri runs the GLM-5.2 MoE model without a GPU by keeping dense weights in RAM and streaming experts from NVMe. Here are the trade-offs.
Xiaomi Released a Frontier AI Model Anonymously—and Developers Chose It
Hunter Alpha was an early MiMo-V2-Pro build. Its uptake shows how price, output quality, and data jurisdiction shape model selection.
Why LLMs Emerged and Their Role in Enterprise AI
How the limits of search and rule-based systems led to LLMs—and why a model is an interface to knowledge, not enterprise memory.
How to Stop an AI from Inventing Numbers in Your Data
A practical Claude Desktop-inspired setup: DuckDB performs calculations while the model works through schemas, tools, and guarded access.
Fine-tuning Qwen 3 0.6B: from 10% to 92% question classification accuracy
A home chatbot experiment: fine-tune a tiny local LLM with Unsloth and two-letter category codes instead of label names for metadata-aware RAG.