AI News Report, October 16: META RELEASES MUSE SPARK OPEN, AN 816B OPEN-WEIGHT MODEL THAT RUNS ON TWO MACS

Meta released Muse Spark Open on Thursday, an open-weight AI model with 816 billion parameters that can be downloaded and run on two 512 GB Macs in its 4-bit 533 GB form, according to The New Stack's AI News Report. The model reads text, images, video and audio, and the report also notes Reflection's 81.6% score on SWE-Bench Verified, Microsoft's MAI-Base-2 with 80 billion active parameters pre-trained 4x to 12x more efficiently than the strongest open base model, and FTC civil investigative demands to AI companies over recent security incidents.

The new US lab's first model has 23B of its 501B parameters active per token and a 1M context, and Reflection reports 81.6% on SWE-Bench Verified. Apache 2.0 has no revenue line, so you can host it for clients. It reads text, images, documents and video with 12B of its 120B parameters active, and NVIDIA claims up to 5x the single-GPU agent throughput of open models its size. Checkpoints, NVFP4 included, are on Hugging Face. Opus 4.8 and Opus 5 users move to Opus 5.5, and GPT-5.6 Terra users to GPT-6.1 Sol or GPT-6 Luna, in every Copilot mode. Check your org's model policies and any pinned model names before then. It flags vulnerabilities in PHP, Shell, Terraform and Dockerfiles, where CodeQL has gaps, but its findings cannot block a merge. It spends Code Security AI Credits, so set a budget before you turn it on. The open models run from 5M to 209M parameters and forecast a day ahead in 10 to 140 ms on an 8-thread server CPU, no GPU needed. Try it on ticket volume or capacity data without any training. The new Forward action is in public beta, and DLP matches can now be checked against your own data, like a live GitHub token, before a block. Traffic sent to your Worker bills as normal Worker requests. His scanner backend finished a 3600 dpi scan in 34.6 seconds against 54.4 for upstream SANE, and low-light camera brightness rose from 31.1 to 90.9. His trace-and-diff method works for other odd hardware. Claude Code on Opus 5.5 scored 22.0 of 23 on the bigger task without comments and 21.7 with them. Read it before you pay an agent to write comments, and note how much single runs vary. Across 5,240 setups, changing only the order of the track descriptions flipped Jev's choice in 370 of 2,620 pairs. Test any decision model on your own data before it touches a workflow. The excerpt walks four disaster recovery patterns and suggests a backup model vendor, noting Uber reportedly spent its yearly AI budget in four months. Read it if your help desk runs on one model. While enterprise leaders overwhelmingly believe in the promise of synthetic data, adoption lags. As a result, synthetic data has great, The post Only… · The New Stack The civil investigative demands work like subpoenas and ask for records on recent security incidents and safeguards, Axios reported. The FTC also plans to call executives from both companies to testify. The critical advisory says agents used exposed credentials and loose cloud roles, and none of the companies noticed. It says to rotate exposed secrets, enforce MFA and keep logs for at least 180 days. Agents on GPT-5.6 Sol chained three Chrome bugs for about $40,000 in AI charges, and Google paid a $250,000 reward. Check that every managed Chrome is on 152.0.7977.82 or later and restarted. The draft lists 40 strategic technologies, 23 tied to AI, including chips and large language models, with comments open until November 15. If you license Chinese models or chips, watch the registration rule. Proofpoint pitched the deal as security for AI agents that can reach too much company data, and it should close in the first half of 2027. Expect data-access controls to come bundled with email security. MAI-Base-2 has 80 billion active parameters, and Microsoft says it was 4x to 12x more efficient to pre-train than the strongest open base model. Its weights go to trusted partners through Foundry. Use above that is billed to a project code, after some staff spent as much as $1,000 a month, 404 Media reports. If you buy AI seats for a team, budget by usage. It is a letter of intent with no price, and AISLE says its team found over 500 unknown CVEs in about a year. Faster bug finding means more patches landing on your desk. The BriefMeta released Muse Spark Open on Thursday, an open-weight AI model, which means anyone can download the trained model and run it on their own machines. It has 816 billion parameters, reads text, images, video and audio, and its 4-bit version 533 GB fits on two 512 GB Mac Studios or four high-end Nvidia GPUs, or you can rent it on OpenRouter for $1.25 per million input tokens. The license is free for nearly everyone, but products must show "Built with Muse Spark Open", and a company with more than $20 million in yearly revenue that sells access to it through an API needs a separate deal with Meta first. Level UpTest Muse Spark Open on one real task this week before you trust the benchmarks. The quickest route is an API: OpenRouter lists it at $1.25 per million input tokens and $4.25 per million output tokens, and Meta recommends a reasoning effort of high or xhigh for coding and agent work. If you ship it in a product, add the line Built with Muse Spark Open to your docs or settings page, which the license requires. →Meta: Build with Muse Spark Open deploy guide