Liquid AI's open-weight LFM2.5 ships + Groq raises 650 million dollars

Liquid AI released LFM2.5-8B-A1B, a mixture-of-experts model with roughly 1 billion active parameters out of 8 billion total, pretrained on 38 trillion tokens and post-trained to a 128,000-token context window. The release is open-weight with no use restrictions and ships day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX runtimes. Liquid reports around 253 tokens per second on an Apple M5 Max, 146 on a Ryzen AI Max Plus, and roughly 30 tokens per second on flagship smartphones — putting credible laptop and phone inference into the same parameter band as the closed flagship-mini tier.

Groq is raising 650 million dollars to refocus the company on its inference neocloud — the on-demand cloud platform powered by Groq's own AI chips — after a December 2025 arrangement saw Nvidia pay 20 billion dollars for senior Groq engineering talent and a hardware license. Existing investors are leading the round, with Disruptive and Infinitium committed to fill any unsubscribed shares. Adam Winter is now interim CEO and Matt Eng interim CFO. The company argues inference is a much bigger market than training right now, and the wedge for the smaller team that remains.

At the AI Now Summit, Mistral named Airbus (commercial aircraft, helicopters, defense, and space), BMW Group (the carmaker's Large Industry Model initiative for multimodal reasoning on engineering data), and ASML (semiconductor parts design and surrogate modeling) as production customers for its new Mistral Physics AI line. The lab also confirmed a new 10 megawatt data center in Les Ulis south of Paris, dedicated to inference and scheduled to open in the third quarter of 2026, alongside an existing 40 megawatt Paris site and a planned Swedish facility — Mistral's answer to the inference supply-chain risk that has loomed over European AI customers.

Following Cognition's 1 billion dollar raise at a 26 billion dollar valuation announced May 27, CEO Scott Wu publicly reframed Devin's positioning as augmentation rather than replacement: "We've never thought about it as replacing humans. We are all programmers ourselves." Wu disclosed that 89 percent of Cognition's own engineering output is committed by Devin — with Windsurf, the acquired coding competitor, also contributing. He characterized Devin's current capability as somewhere between a junior and a mid-level engineer depending on task complexity, deliberately avoiding higher-level replacement claims.

A widely shared engineering post from Quandri by backend engineer Chloe Kim argues that the Model Context Protocol is over-rated for most developer workflows. Kim instrumented their stack of MCP servers and found they consumed roughly 21,000 tokens — about 10.5 percent of Claude's 200,000-token context window — even when most tools sat unused. A single Linear-issue lookup ran around 12,957 tokens via MCP versus 200 tokens via direct command-line equivalents, a 65-times difference. Kim's recommendation is a Skills-style on-demand loading pattern paired with command-line-first integrations, reserving MCP for services without CLIs or where shared team authentication is essential.