AI news6
Meta releases Muse Spark Open weights for local and hosted use
Meta released Muse Spark Open on October 15, with 816 billion total parameters, 42 billion active parameters and a one-million-token context window. The weights are downloadable under the Muse Model License with limited commercial terms. Meta demonstrates local use across two M5 Ultra Mac Studios, while several providers offer hosted access.
research.meta.ai ↗Why it matters
The new weight release adds a local deployment option, a separate development from the Muse Wonder models covered earlier this week.
Microsoft details a 2.8-trillion-parameter model still in post-training
Microsoft’s October 15 technical report describes MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active parameters. It trained on 16,000 GB200 GPUs over 38 days. Microsoft reports stronger held-out pre-training results than the open-weight models it tested, but the model is still undergoing post-training and has not been released to customers.
microsoft.ai ↗Why it matters
This new training report shows Microsoft developing its own model pipeline, distinct from its weather software and earlier decision-model release.
Liquid launches a multimodal decision API at US$0.042 per million input tokens
Liquid AI released d1-ultra on October 15 for decisions based on text, images or both. It returns probabilities for yes/no, label-selection and scoring questions without generating output tokens. The Liquid API charges US$0.042 per million input tokens, with each question billed as a separate prompt; Vercel and OpenRouter availability is forthcoming.
liquid.ai ↗Why it matters
This new multimodal probability service addresses structured decisions rather than the local reasoning model covered on Thursday.
NVIDIA releases a 120-billion-parameter visual model for specialised agents
NVIDIA’s October 15 release introduces Nemotron 3.5 Super VL, with 120 billion total parameters and 12 billion active per token. Its open weights combine text, image, document and video understanding with reasoning and tool use. NVIDIA says Ramp is adapting it to turn receipts into structured data and expects up to 50% higher throughput and tenfold lower latency; those are projections.
developer.nvidia.com ↗Why it matters
The newly released visual model adds a different capability from the PAIR software and forthcoming Bolt model described on Wednesday.
Black Forest Labs releases a 5.6-billion-parameter image model
Black Forest Labs released FLUX 3 Image Klein on October 15 for image generation, editing and combining up to ten reference images. The open-weight release includes base and distilled versions under the FLUX Kommunity License. The developer says its INT8 version runs through ComfyUI or Diffusers on NVIDIA GPUs with as little as 8GB of video memory.
bfl.ai ↗Why it matters
This new, smaller Klein release creates a local editing option, distinct from the older full FLUX 3 rollout previously rejected as stale.
Perplexity open-sources the sandbox behind its Computer agent
Perplexity announced the open-source release of SPACE on October 15, after deploying the system internally in July. Each sandbox uses a small virtual machine, with controls for network access and credentials. The infrastructure supports isolated code execution, pausing, recovery and session branching.
perplexity.ai ↗Why it matters
Publishing the implementation now gives developers code to inspect and adapt, rather than relying solely on a model to respect access limits.
AI in the nonprofit sector3
Six METR researchers gain access inside Anthropic for independent evaluation
Anthropic said on October 15 that it has onboarded six employees of the nonprofit METR with company hardware, office space and access comparable to its internal risk teams. METR can publish its assessments without an Anthropic veto. Access excludes defined categories, including customer data, privileged legal information and model architecture, and Anthropic can require redactions within those categories.
anthropic.com ↗Why it matters
The new access arrangement gives nonprofit evaluators more evidence than an external model test, while its publication and redaction rules define the limits of that oversight.
New nonprofit commits US$25M to AI and cybersecurity support for public institutions
Anne Neuberger announced Securing America’s Infrastructure Foundation on October 15, with US$25 million committed by a16z’s founding partners. The new 501(c)(3) plans to deliver software, training and cyber incident response at no cost, beginning with a two-city pilot. It aims to reach roughly five to ten anchor cities over five years, with small municipalities, school districts and rural hospitals among its intended beneficiaries.
a16z.com ↗Why it matters
The proposed support combines implementation staff and tools for institutions with limited technical capacity; it is a service programme rather than an open cash-grant call.
Open Intelligence Foundation opens a ten-day community AI programme
The nonprofit Open Intelligence Foundation is stewarding Open Source AI Week across the San Francisco Bay Area from October 16 to 25. Its October 16 announcement lists more than 60 community-run events, including an on-device AI hackathon and open-model safety and policy debates. Individual events require registration.
globenewswire.com ↗Why it matters
The programme creates shared places for civil society, researchers and developers to examine and build open AI systems, with the event week now beginning.