Blog

My hands-on StepFun Step 5 Preview review: a 600B/27B MoE reasoning model at $1/$2.70 per 1M tokens. The pricing is real, the benchmarks are self-graded. Here is my verdict.

StepFun Step 5 Preview is a 600B/27B sparse MoE model with a 1M-token context, open weights due Oct 15. Here are the specs, pricing, benchmarks, and my take.

A hands-on Qwen Image 2.1 review: what the 7B open-weight model does well, the native transparency trick, and the license catch nobody's mentioning.

An honest Laya AI review: the open-source 421M decision model is genuinely fast and calibrated, but the headline accuracy hides a fine-tuning caveat. Here is where it wins, where it does not, and who should use it.

Laya AI is Convai's open-source, 421M System 1 decision model: 33ms typed answers with calibrated confidence, no text generation, and an honest zero-shot caveat.

TypeSafe calls Jev an ultrafast System One model at 70-500ms a decision. Here is what the speed claim really means, where it holds up, and where it does not.

TypeSafe Jev pricing broken down: $0.042 per million input tokens, output free, no plan tiers yet, and what a System One model actually costs to run in production.

A hands-on TypeSafe Jev review: what the System One model actually does, whether the speed, price and 'can't hallucinate' claims hold, and where a typed-decision model fits real work.

A plain-English guide to TypeSafe Jev, the System One model that turns unstructured state into typed yes/no, pick-one, and score decisions your code can trust.