ModelsVision-Language Models: What to Test Before You Commit
Native image input is now common in open-weight models, but published multimodal benchmarks predict almost nothing. Three probes that separate them.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
Picking between models — and testing them on your own work.
Models
ModelsWhen to Switch Models: A Threshold Worth Crossing
New models ship constantly and switching is never free. The signals that justify a migration, the ones that do not, and how to run the decision.
Read
ModelsWhich Open Model Fits Your GPU: A Memory-First Guide
Total parameters set the memory floor, active parameters set the speed. How to work out what your hardware can actually serve before reading a benchmark.
Read
ModelsWhy Benchmarks Disagree, and Which One to Believe
Two credible evaluations can rank the same models differently without either being wrong. How to read that disagreement instead of picking the flattering one.
Read
ModelsBest Model for Code Review: Optimise for Precision, Not Recall
An automated reviewer that flags everything gets muted within a week. How to pick and evaluate a model on false-positive rate, and why review has odd economics.
Read
ModelsBest Model for Debugging: Judge It on How It Handles Being Wrong
Debugging is a search problem, so the model that wins is the one that forms testable hypotheses and abandons them fast. How to evaluate that properly.
Read
ModelsBest Model for Test Generation: Coverage Is Not the Metric
Models are very good at producing tests that pass and prove nothing. Use mutation score to pick one, and treat test writing as your cheapest high-volume task.
ReadModelsDeepSeek V4 Pro: Frontier Reasoning at Commodity Prices
A 1.6T MoE that activates 49B per token, ships under MIT, and costs $0.435 per million input tokens. Where V4 Pro is strong, where it is not, and Pro versus Flash.
ReadModelsGLM-5.1 vs GLM-5.2: Same Size, Same Price, Different Job
Z.ai kept the architecture and the price list identical and changed what the model is for. The context jump, the long-horizon gains, and the token cost nobody mentions.
ReadModelsGLM-5.2 vs DeepSeek V4 Pro: Endurance or Raw Problem Solving
Both are MIT-licensed open models, so the licence is a wash. The split is agentic endurance against algorithmic power, at a three-to-five times price gap.
ReadModelsGLM-5.2: The Open Model Built for Long-Horizon Coding
Z.ai shipped a 744B MoE with 40B active, MIT-licensed weights and the first open-weight Terminal-Bench 2.1 score above 80. A technical read on what that means.
ReadModelsHow to Choose a Coding Model Without Trusting Anyone’s Chart
Six properties decide whether a model is good at your codebase — and only one of them shows up on a leaderboard. A practical framework for picking.
ReadShowing 97–108 of 119 articles