New AI models & training tweak

Training a large artificial intelligence model usually means building a big system first and trimming it later; Massachusetts Institute of Technology researchers say they can now shrink some models while training is still underway. (news.mit.edu) The method, called CompreSSM, targets state-space models, a neural-network design used in language, audio, and robotics systems. The team said it can identify low-value internal parts after about 10 percent of training, then discard them so the remaining 90 percent runs like a smaller model. (news.mit.edu) State-space models work like a running memory of what came before, updating an internal state as new tokens or signals arrive. CompreSSM borrows a control-theory measure called Hankel singular values to rank which hidden dimensions matter most before cutting the rest. (news.mit.edu) That efficiency push landed in the same week as three product launches aimed at day-to-day work rather than abstract leaderboard claims. Meta introduced Muse Spark on April 8, 2026; Z.ai published GLM-5.1 on April 7; and Anthropic rolled out Claude Managed Agents in public beta on April 8. (about.fb.com) (z.ai) (anthropic.com) Meta said Muse Spark is the first model from Meta Superintelligence Labs and now powers the Meta AI app and website. The company said the model will roll out in coming weeks to WhatsApp, Instagram, Facebook, Messenger, and artificial-intelligence glasses, with private-preview application programming interface access for selected partners. (about.fb.com) Meta also said Muse Spark is built for “smarter and faster” assistance inside its own products, including image understanding, shopping help, and answers that can eventually cite recommendations and posts shared across Instagram, Facebook, and Threads. The company said new features begin in the United States before expanding to other markets. (about.fb.com) Z.ai framed GLM-5.1 as a model for “long-horizon” agentic engineering, meaning jobs that take many rounds of planning, testing, and revising instead of one answer. In its April 7 post, the company said GLM-5.1 scored 58.4 on SWE-Bench Pro and kept improving through more than 600 iterations and 6,000 tool calls on one vector-database task. (z.ai) Z.ai also said GLM-5.1’s weights are publicly available on Hugging Face and ModelScope, and that the model supports local deployment through vLLM and SGLang. That makes it one of the week’s more accessible releases for developers who want to run or inspect a model outside a closed chat product. (z.ai) (huggingface.co) Anthropic’s launch was less about a new base model than about the plumbing around one. Its documentation says developers can choose between the Messages Application Programming Interface for direct prompting and Claude Managed Agents for long-running, asynchronous work on managed infrastructure. (docs.claude.com) (anthropic.com) Anthropic said Managed Agents package the session log, the harness that routes tool calls, and the sandbox where code runs into a hosted service designed to stay stable as underlying implementations change. The company’s Agent Software Development Kit separately exposes built-in tools for reading files, editing code, running shell commands, and searching codebases. (anthropic.com) (docs.claude.com) Put together, the week’s releases pointed at the same bottleneck from two sides: training still costs too much, and useful agents still take too much engineering glue. The new pitch from labs and platform companies is not just stronger models, but models and tooling that spend fewer compute cycles and fewer developer hours getting real work done. (news.mit.edu) (z.ai) (docs.claude.com)