Meta released Muse Spark Open, an open-weights model, and also rolled out closed models Muse Wonder and Muse Wonder Full, alongside mentions of other labs’ scores and upcoming weight releases. Muse Spark Open is an 816B total / 42B active parameters model with a 1M-token context window. As a result, the gap between top US and Chinese open-weight models is described as roughly halving, with more options for consumers and a more competitive Western frontier landscape.
Meta’s chief AI officer Alexandr Wang built and scaled the Muse personal AI agent after Mark Zuckerberg sought his advice, culminating in the rollout of Muse Wonder as a “frontier-class” model. Meta expects 2026 capital expenditures of $130 billion to $145 billion, nearly double last year’s $72.2 billion. The effort is driven by recruiting and internal rebuilding focused on getting a model that performs well enough to justify the larger AI spending.
Muse released its Muse Spark Open and Reflection Beam models alongside multiple frontier and decision-model updates covered in this AINews roundup. Muse Spark Open is an 816B-total, 42B-active MoE with weights quantized so experts-only fit to about 530GB. As a result, open-weight licensing, pricing, and benchmark rankings shift across cost, agent performance, and inference hardware/serving notes.
TypeSafe launched Jev, a decision-model that outputs labels/scores or yes/no probabilities without generating text, and within about 4 weeks at least 20 other Jev alternatives were shipped by companies and developers. Cloudflare’s Clef (27B) scored 94.20 on BANKING77 F1 versus Jev’s 79.74 in Cloudflare’s own run. Developers now have many more hosted and self-hostable decision-model options, with differences in licenses, hardware support, pricing, and some models claiming Jev SDK/API compatibility.
StepFun released Step 5 Preview open-weight checkpoints, including FP8 and BF16, on Hugging Face and ModelScope. The model has 600B total parameters with 27B active per token and supports up to a 1M-token context. Developers can now self-host it for agentic coding and long-document/finance workflows, with specified data-center GPU requirements (4x B300 for FP8 or 8x B300 for BF16) instead of using only the API.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
