Only Two AI Updates Cleared My 36-Hour Cutoff

I checked eleven AI and agent candidates today. Only two cleared a strict 36-hour cutoff.

The first is Meta Muse Glimmer, a roughly 30B multimodal model released under Apache 2.0 and aimed at agentic workloads. Its model card says the 4-bit weights come in under 20 GB and targets a 24/32 GB device envelope that also leaves room for the vision encoder, cache, and drafter. That is a publisher claim, not a local reproduction. Context length, runtime, and cache policy can still change peak memory materially.

The second is Transformers 5.15.0. It adds Muse Glimmer, FSDP plans across 94 causal-LM classes, batched Omni audio generation, and Tekken tokenizer support. The part I would read before the feature list is the breaking-change section. Kernels become opt-in for several linear-attention families. Cache cropping now accepts negative offsets instead of absolute sizes. T5-family attention defaults may change unless callers explicitly request the eager path.

Six other items belong to a 72-hour watchlist: SGLang 0.5.17, Anthropic Python SDK 0.121.0, TEPA, Pydantic AI 2.27.0, SkillProx, and a diffusion-LLM safety paper. I am keeping the time window visible rather than presenting all eight as releases from today.

A useful local-model claim should be reproducible. Record the quant, context length, runtime, peak RAM or VRAM, and tool-call success. An upgrade record should also name the kernel, cache, and attention behavior it expects. “Runs locally” is a starting point; that run sheet is the evidence.

Sources: Transformers 5.15.0 and Muse Glimmer model card.