Meta Goes Closed, Google Goes Apache, and a 26-Person Startup Goes Frontier

News

Meta Breaks From Llama Lineage With Closed-Source Muse Spark

After years of being open-source AI's biggest champion, Meta launched Muse Spark as a proprietary model, marking a sharp strategic reversal. The model is built by Meta Superintelligence Labs under Alexandr Wang and claims competitive performance with GPT-5.4 and Claude Sonnet 4.6, but as thenextweb.com notes, "Muse Spark breaks that pattern" set by Llama. If you were building on Llama expecting Meta to stay open forever, this is your signal that the bet has an expiry date.

Tools

Gemma 4 Lands Under Apache 2.0: Google's Most Capable Open Models Yet

Google DeepMind released Gemma 4 on April 2 with four model sizes from a 2B edge variant up to a 31B dense model, all under a fully permissive Apache 2.0 license. The 31B scores 89.2% on AIME math and 80% on LiveCodeBench, benchmarks that would have been frontier-only numbers a year ago. For builders who want a truly open, commercially usable model that runs offline on consumer hardware, this is the one to pull from blog.google.

Gemma 4 Under Apache 2.0 Is the First in the Series Without Licensing Friction

Prior Gemma versions shipped with a custom Google license that created commercial gray areas for enterprise teams. Gemma 4 is the first to use OSI-approved Apache 2.0, and it matters because it now competes on equal licensing footing with Llama and Qwen. The 26B MoE variant activates only ~4B parameters during inference, meaning it can run on a 16GB consumer GPU, per opensource.googleblog.com.

Analysis

Arcee Trinity-Large-Thinking Fills the Void Meta Left in Open-Weight AI

With Meta now going proprietary and Chinese labs pulling back from public releases, Arcee is one of the few credible Western options for truly open frontier-scale weights. Trinity-Large-Thinking runs on OpenClaw and has become a drop-in for developers who had been using Claude but lost subscription access to third-party tools. techcrunch.com notes all Trinity models carry "the gold standard for OS licenses, Apache 2.0," which is the key differentiator from Llama's restricted license.

Muse Spark's "Thought Compression" Training Trick Is Worth Watching

Meta claims Muse Spark achieves competitive reasoning using more than ten times less compute than Llama 4 Maverick by penalizing the model during reinforcement learning for excessive reasoning token use. That training efficiency story, not the benchmark numbers, is the technically interesting part here. According to cnbc.com, this enabled "smaller AI models that are as capable as its older midsize Llama 4 variant for 'an order of magnitude less compute.'"

Arcee Trinity Technical Deep Dive: SMEBU, Muon Optimizer, and 4-of-256 Expert Routing

Arcee used a novel load-balancing technique called SMEBU (Soft-clamped Momentum Expert Bias Updates) to prevent expert collapse at extreme sparsity, plus the Muon optimizer during a 17T-token pretraining run. Only 4 of 256 experts activate per token, which is what makes this 400B model fast enough to be practical on real hardware. marktechpost.com has the full breakdown for anyone doing architecture-level evaluation.