Stripe acquires model gateway OpenRouter as DeepSeek and Liquid AI expand open-weight tooling
The infrastructure for open models is consolidating and expanding; Stripe is buying API gateway OpenRouter in a multibillion-dollar deal, while DeepSeek open-sources its agent runtime and Liquid AI releases quantized models.
Artificial Intelligence··Evening
OpenRouter acquisition
Payments company Stripe is acquiring OpenRouter, a gateway that spreads developer requests across different AI models, for a reported 7.5 billion dollars. The founders will take about 1.5 billion dollars and investors the remaining 6 billion dollars, while the gateway is expected to keep operating independently. The startup was valued at 1.3 billion dollars in May 2026. The Collison brothers' investor letter tied the purchase to the platform's broad utility, noting that Stripe counts as one of the world's largest developer platforms and the integration serves that audience.[1]
DeepSeek agent framework
DeepSeek released DeepSeek Harness, an open-source execution runtime for building autonomous AI agents, under the MIT licence. The preview version ships with four runtime modes: Standard, Code, Minimal and Creator. The system uses a micro-kernel design where model adapters, tool registries, sandboxes and user interfaces load as independent plugins, allowing developers to switch models without touching the core. Configuration is written in YAML or JSON, and an append-only event log records each user message, tool call, intermediate reasoning state, token metric and sub-agent dispatch.[2]
Liquid AI quantized models
Liquid AI published Q4_0 builds of four LFM2.5 models trained with quantization-aware distillation, a method that distils a high-precision teacher into a quantized student. The variants include LFM2.5-230M and LFM2.5-2.6B, all offered in GGUF format. The company reports that the quantized builds recover about 97 per cent of what the original models score across multiple benchmarks. The measurements were run on evaluations including GPQA Diamond, MMLU-Pro, IFEval, IFBench and GSM8K along with AIME25.[3]
