Meta AI & Llama news & updates
One ongoing page for Meta AI, Muse, Llama and Meta model research. Verified developments are added to the timeline below, with the sources and context needed to follow what changes.
No material news since the previous check. Coverage: 17 October 2026 at 12:08 BST–18 October 2026 at 12:09 BST. All historical entries and existing visuals remain on this page. Next scheduled scan: 19 October 2026 at 12:00 BST · Europe/London.
The ongoing timeline
Newest update firstHistory stays here
Meta releases gated Muse Spark Open multimodal weights
Meta released gated BF16 and NVFP4 weights for Muse Spark Open, a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Its model card says self-hosting needs at least 700 GB of VRAM for NVFP4 or 2 TB for BF16. The Muse Model License requires attribution and imposes separate terms on larger model-as-a-service businesses, so this is open-weight rather than permissively licensed open-source software. Meta recommends deployment-specific guardrails, and its benchmark and performance claims remain vendor measurements.
Why it matters: The release makes frontier-scale multimodal weights downloadable, but gated access, exceptional memory requirements and a restricted commercial licence narrow who can deploy them independently.
Meta’s article and model-card release are dated 15 October 2026. Hugging Face records showed the BF16 repository updated at 16:46 BST, its collection at 17:54 BST and the NVFP4 repository at 18:18 BST, all after the prior tracker check; these repository timestamps establish availability rather than a precise article-publication time.
| Question | What the model cards say |
|---|---|
| Can I download immediately? | Repositories are gated; approval and licence terms apply |
| NVFP4 deployment | At least 700 GB VRAM, according to Meta |
| BF16 deployment | At least 2 TB VRAM, according to Meta |
Meta publishes ten manuscripts from a human-led Muse health-data experiment
Meta described a four-week research exercise in which medical researchers used an internal scaffold powered by Muse Spark 1.3 and Muse Wonder. The models assisted with hypotheses, analysis code and draft manuscripts using public data and a licensed database of five million de-identified health records, while researchers retained responsibility for decisions and conclusions. Meta published ten manuscripts after cohort feedback, biostatistical checks and contracted outside review. These are exploratory retrospective studies rather than clinical trials or validated medical guidance; the published reviews identify limitations including possible surveillance bias.
Why it matters: The project offers a concrete record of supervised research-agent use, without showing that faster manuscript production establishes clinical validity.
Manuscripts and company account published 14 October 2026; officially shared that evening. The reported research exercise lasted four weeks.
Senate Finance Democrats ask Meta to explain AI data-centre research-tax claims
Seven Democratic senators publicly asked Meta to explain whether federal research credits helped fund AI data-centre construction and operations. Their letter seeks annual amounts and the rationale for treating facilities as experimental pilot models, with a reply requested by 31 October. The senators cite about $6.6 billion in Meta research credits over 2023–2025, but that total does not establish how much was tied to data centres. This is an oversight inquiry, not a committee finding, an IRS ruling or proof that the company broke tax law.
Why it matters: The request scrutinises how tax support for research relates to the scale of AI infrastructure spending, while leaving the factual and legal questions open.
The original letter is dated 13 October 2026; its public announcement and Bloomberg Government coverage are dated 14 October, with the report at 19:04 BST.
Facebook begins rolling out natural-language controls for Feed and Reels
Meta announced Your Algorithm, a Facebook control for adding and managing the topics people want to see in Feed and Reels. Users can also explain what they like or dislike after selecting Interested or Not interested on a post. Meta says large language models interpret that freeform feedback alongside existing recommendation signals and adapt rankings within the same session. The announcement does not identify the model or give a region-by-region completion schedule, so this is a rollout rather than proof that every account already has it.
Why it matters: The update lets people give explicit instructions to a recommendation system, rather than relying only on inferred interests; its effectiveness remains Meta’s claim.
Rollout announced 13 October 2026; official publication: 17:00:29 BST.
European privacy board issues preliminary guidance for smart glasses
The European Data Protection Board adopted a statement on smart glasses, including devices with AI translation or identification features. It says manufacturers and linked-service providers retain responsibility for their own processing and calls for reliable, conspicuous recording indicators. The statement warns that relying on legitimate interests for further commercial reuse or AI training is particularly difficult to justify when bystanders are unaware of collection. This is sector-wide preliminary guidance, not a Meta-specific infringement finding, fine or new product ban.
Why it matters: The guidance is relevant to AI-enabled eyewear and its surrounding services, while leaving individual processing arrangements subject to case-by-case assessment.
Statement adopted 13 October 2026; EDPB announcement published at 13:00 BST. No precise adoption time is recorded.
Meta releases Muse Wonder models with different API and coding access
Meta announced Muse Wonder and Muse Wonder Full on 12 October, with multimodal input, text output and a context window of about one million tokens. Wonder is available through the Meta Model API public preview and Muse Code, with published API rates of US$4 per million input tokens and US$20 per million output tokens; Full is restricted to Muse Code’s Power Usage and Maximum Usage subscriptions. These are hosted offerings rather than new open Llama weights, and integration into the consumer Muse app remains forthcoming. Muse Code’s Contributor mode is on by default for subscribers and permits training use of submitted content, but it can be disabled; Standard mode does not use that content for training.
Why it matters: The release expands Meta’s developer-facing model options, but variant access and data-use settings matter more than treating company benchmark claims as an independent ranking.
| Variant | Meta Model API | Muse Code |
|---|---|---|
| Muse Wonder | Public preview | Pay-as-you-go and every subscription plan |
| Muse Wonder Full | Not available | Power Usage or Maximum Usage subscription |
Availability checked on 13 October. Selecting maximum reasoning on Wonder does not select Wonder Full. See the linked official model documentation for account requirements.
Model release / announcement: 12 October 2026; availability and data-use documentation checked on 13 October.
Meta expands model safety requirements to training and evaluation
Meta updated its model safety framework to extend containment requirements into training and evaluation, including vetted sandboxes, secure logs and automated monitoring for higher-risk runs. It also clarified considerations for open-weight releases and said it plans a new AI committee of its board. These are Meta’s stated policies and governance plans, not evidence of an independent audit or a new Llama model release.
Why it matters: The requirements influence how Meta develops and releases capable AI models, including potential open-weight models.
Muse expands into small-business workflows
Meta introduced Muse for Small Business with additional skills and connectors for business applications, Facebook Pages and Instagram business accounts. The announcement says publishing, sending and spending require user approval, and identifies Muse as available in the US and Canada. Meta describes free usage for many needs and subscription plans for additional usage, with available connectors listed in the app.
Why it matters: The update extends Meta’s personal agent into business administration and marketing tasks.