AI progress — 16 October 2026

Meta released Muse Spark Open, a coding-optimized open-weight model that it launched together with Apple to run locally on Apple silicon. It is already listed in Amazon SageMaker JumpStart, so it can be deployed on a local Mac or in an AWS account on day one. Elsewhere, a report says the FTC is looking at how Anthropic and OpenAI handle security incidents, and Anthropic has given METR employee-level access.

New models and papers

• Muse Spark Open — Meta’s coding-optimized open-weight model, launched with Apple and built to run locally on Apple silicon. Why it matters: a coding model with open weights that targets a laptop is now a deployment option without a hosted API; AWS lists it in SageMaker JumpStart as well.
• EdgeWise: long-horizon agents with selective cloud delegation — a paper proposing that a local model handles most agent steps and calls a cloud frontier model only for key ones; the authors report that using the large model on a small subset of steps can reach comparable performance. Why it matters: if the result holds outside the paper, it points to a way to cut inference cost on long agent runs. It is a single paper, not yet independently checked.

New tools and software

• llama.cpp v0.7.0 — adds LiliCorr speculative decoding, MTP for GLM5-Next, and a --moe-cache-mib host-memory cache for MoE experts, plus support for a batch of new models including MiniCPM-V 4.7 and Nemotron Super 3.5 VL. Why it matters: the release notes cite up to ~2.2x decode speedup when MoE experts are offloaded, which matters for anyone running large MoE models on limited GPU memory.
• GitHub AI Scan for pull requests — now generally available; it looks for vulnerabilities in pull-request changes and complements CodeQL for languages and frameworks CodeQL does not cover. Why it matters: teams with code in languages CodeQL skips can get PR-time security scanning from GitHub instead of adding another tool.
• Claude Code v2.1.301 — adds CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS, so unattended retry mode ends a request at a usage limit that resets too far away instead of waiting. Why it matters: scheduled and headless runs no longer hang for hours on a distant limit reset; they fail and can be handled.

Overall digest

• Anthropic gives METR employee-level access — METR will publish its own assessments of Anthropic’s risks, incidents and practices, and Anthropic has given up the right to block them, per TechCrunch. Why it matters: it is a concrete arrangement for outside evaluation with publication outside the company’s control; it is one company’s agreement, not an industry rule.
• FTC reportedly probes Anthropic and OpenAI over security incidents — Ars Technica, citing a report, says the FTC may fine AI firms on par with companies that delay data breach reports. Why it matters: if confirmed, incident disclosure timing at AI labs would be treated like breach reporting. This is a report of a probe, not an announced action.
• ChatGPT can see your open Chrome tabs — the ChatGPT Chrome extension can now view the tabs you already have open when you tag Chrome in the ChatGPT web or desktop app; the Atlas browser is gone. Why it matters: the assistant now works against your logged-in browser state, which changes what you are granting when you install the extension.

Quiet day for papers: only one made the cut.