8 articles tagged
Poolside's 118B open-weight coding model is 100x cheaper on input than Claude Fable 5 and scores 70.2% to Fable's 88% on Terminal-Bench 2.1. Here is the routing decision that follows from those two numbers.
Microsoft put Mistral's open-weight Medium 3.5 into Foundry and Azure Local on July 21, 2026, from cloud to fully disconnected. How regulated enterprises should choose a deployment mode, and what control really buys.
DeepSeek V4 is both a cheap API and an open-weight model you can self-host. Here is the 2026 cost math: Flash at $0.14/$0.28 per 1M tokens, H100/H200 rental rates, and the GPU utilization point where self-hosting pays.
Thinking Machines shipped Inkling, a 975B open-weights model with 41B active, 1M context and native audio and vision. Here are the benchmarks against Kimi, DeepSeek and the closed frontier, the real cost, and who should
Kimi K3 leads the Frontend Code arena and ships open weights by July 27, 2026. A senior engineer's read on its coding benchmarks, $3/$15 API pricing, self-host reality and the adopt-vs-wait decision.
GitHub made Kimi K2.7 Code generally available in Copilot on 1 July 2026, the first open-weight model in the picker. It is the cheapest Versatile model on the price list, and Moonshot's own benchmark table puts it
Z.ai's GLM-5.2 is the strongest open-weight coding model of 2026 and a third of Claude Opus 4.8's input price. Self-hosting it is another bill: 753B parameters, 1.5 TB of weights, a GPU node rented by the hour.
US firms routed up to 46% of their OpenRouter tokens to Chinese open models in 2026 because they cost 60-90% less. Here is how DeepSeek, Qwen and GLM compare with OpenAI and Google on price, quality and risk.
