AI progress — 29 October 2026

Google announced Gemini 4 Carbon, its latest frontier model, with claimed gains in coding, agentic tasks and cybersecurity. Ars Technica reports it is not yet available to use and describes it as a modest step over Argon. Elsewhere, llama.cpp and Ollama both shipped support for Liquid AI’s LFM3 family, and GitHub made Copilot Fusion generally available.

New models and papers

• Gemini 4 Carbon — Google DeepMind’s new frontier model; The Verge reports Google claims significant gains in coding, agentic tasks and cybersecurity. Why it matters: Ars Technica says it cannot be used yet, so the claims are Google’s and untested by outsiders for now.
• NVIDIA Nemotron 3.5 Transcribe and Granary v2 — a multilingual transcription model that, per NVIDIA’s post title, can be prompted and adapted, released with a new Granary dataset version. Why it matters: anyone building speech pipelines gets a promptable alternative to fixed-behaviour transcription models.
• TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving — a paper on scheduling agent workloads that alternate between GPU inference and CPU-side tool calls; the authors say existing systems ignore tool-return readiness and use heuristic placement. Why it matters: it targets idle GPU time in agent serving, though the results are the authors’ own and not yet independently checked.

New tools and software

• llama.cpp v0.8.0 — adds the LFM3 family (text, vision, audio, omni), an experimental streaming-audio API, and /v1/realtime and /tts server endpoints plus several new architectures including DeepSeek V4.1 Flash. Ollama v0.40.13 separately brought LFM3 to Linux and Windows. Why it matters: local servers can now do realtime audio in and out through an OpenAI-style endpoint, and LFM3 runs on both main local runtimes.
• GitHub Copilot Fusion — now generally available; Copilot routes and orchestrates which models handle a task instead of the user picking one. Why it matters: teams on Copilot no longer have to choose a model per task, at the cost of handing that choice to GitHub.
• Security extension for the GitHub Copilot app — in public preview, it lets you work through a repository’s code scanning alerts inside the Copilot app. Why it matters: alert triage no longer requires leaving the app where the fixes are written.

Overall digest

• Microsoft recorded $10 billion in revenue from OpenAI — Microsoft says commercial arrangements with OpenAI, including revenue-sharing payments, brought in $10 billion in the quarter ending 30 September. Why it matters: it puts a dollar figure on how financially tied the two companies remain.
• Google expands AI Search ads globally — announced alongside Alphabet’s Q3 earnings, which TechCrunch says showed 25% revenue growth while core ad growth slowed. Why it matters: advertisers and publishers will now see ads inside AI search results outside the initial markets.
• Google AI gets cagey when asked who won the 2020 elections — Ars Technica reports Google’s AI answers hedge on the question in the US but are more forthright outside it. Why it matters: it is a concrete case of a deployed model’s answers varying by region on a settled factual question.