BOTS FIRED

TUESDAY, OCTOBER 13, 2026 | READ TIME: ~8 MIN

Google and Meta expand model access while Liquid targets agents on devices.

The models are arriving through APIs, open weights and local context layers. Microsoft supplies the next hardware roadmap, while NVIDIA opens more of the machinery that keeps GPUs busy. Plenty to test before handing over the keys.

🤖 FRONTIER MODELS

Google opens Gemini 4 Argon to developers and subscribers

Google made Gemini 4 Argon broadly available on October 13, turning a model announced to limited users in September into a product developers and subscribers can use. Paid Gemini API customers can call gemini-4-argon, while Google AI Pro and Ultra subscribers can access it in the Gemini app. Google positions Argon for complex coding, research, and long-running agent tasks. Its API uses one shared 1,048,576-token budget for input, reasoning, and output. Streaming calls return up to 65,536 tokens per request; non-streaming calls cap at 32,768. Current SDKs automatically continue longer generations across requests. Introductory API pricing runs through November 30 at $2 per million input tokens and $10 per million output tokens, including reasoning; the published API rates double on December 1. The release makes a frontier model available to builders with a documented endpoint and price. Its performance claims remain Google's, and teams will still need to measure reliability, latency, and cost on their own workloads.

Meta releases Muse Wonder models for sustained agent work

Meta launched Muse Wonder and Muse Wonder Full on October 12, its newest models built for personal and professional agent work. Wonder is available to developers through the Meta Model API and Muse Code; the Full version is available to Muse Code subscribers. Meta says the models can work through long tasks such as interpreting an insurance claim or lease, handling expert material, and building software. Two features define the release: Active Memory carries relevant information across conversations, and Active Focus is meant to keep an agent working through errors and alternate approaches while asking for clarification or approval when needed. Wonder is also slated to reach Meta's Muse product, initially for subscribers. The release emphasizes assistants that retain context and pursue work over time. Meta describes the models as state of the art, but that is its own assessment; a fuller safety evaluation report had not yet been released with Meta’s model announcement.

Liquid AI releases 24B open-weight model for local agents

Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model built for local agents. Its mixture-of-experts design has 24 billion total parameters but activates about 2 billion per token. The company extended context from 32,768 to 131,072 tokens and offers base and post-trained weights, plus support for llama.cpp, MLX, vLLM, SGLang and ONNX. Liquid reports 184 output tokens per second on an Apple M5 Max and says the model leads or matches a larger rival on five of eight tool-use benchmarks. Those are company-run measurements under specified conditions, not independent proof of performance on every device. The same results show weaker scores on WebArena and SWE-Bench Verified than the larger comparison model, which matters for browser and coding agents. The weights let developers test those trade-offs locally rather than relying only on the launch slides. Developers can inspect the Liquid AI model release and benchmark table before choosing a workload.

🦾 AGENTIC AI

Sentient OS pairs local context with computer-use agents

Sentient OS announced version 2.0 for macOS on October 12, positioning an on-device language model as a personal context layer for computer-use agents. The app reads sources a user connects, including files, messages, notes, and calendars, then builds an editable Markdown knowledge base on the Mac during overnight processing. That context feeds Sidekick, which can operate apps through its own cursor, and Double Tap, which drafts replies from the current field. The company says the app and supporting infrastructure are open source under AGPL. Sentient’s architecture description describes a Gemma 4 E4B model for local understanding and a separate user-selected frontier model for more demanding work; users can choose a hosted subscription or compatible local endpoint. The release is a concrete example of agents combining private local indexing with action in everyday software. Double Tap’s built-in model sends screen and knowledge context to cloud providers; fully local operation requires choosing a local model.

⚙️ HARDWARE & CHIPS

Microsoft unveils Maia 300 with production deployment planned for 2027

Microsoft unveiled Maia 300 on October 12 and said its next AI inference system will begin production deployment in 2027. The company plans to scale the broader Azure Maia family to multi-gigawatt installations from that year, with Maia becoming its primary inference system for products and services. Maia 200 is already handling some Microsoft Copilot production work. A full Maia 300 rack is designed around 96 accelerators and a 167-kilowatt power envelope, with 27.6 terabytes of high-bandwidth memory and 1.7 petabytes per second of aggregate memory bandwidth. Microsoft claims that configuration can deliver up to seven times Maia 200's rack-level token throughput and more than 2.5 times its performance per watt. Microsoft’s architecture announcement shows how tightly hyperscalers are coupling chips, memory, networking, and software for long-context agents. Those performance figures are Microsoft-reported, and the 2027 production target remains ahead; this is an infrastructure roadmap with concrete specifications, not a live Maia 300 deployment.

🏢 ENTERPRISE & APPS

Meta updates GEM to train one ads model across its platforms

Meta has rebuilt its Generative Ads Model to learn across its advertising platforms rather than optimizing each surface separately. The company introduced GEM in 2025; its October 13 update describes a new Kunlun architecture, longer histories of user activity and one central model that predicts clicks and conversions across Facebook, Instagram and other placements. Meta says downstream models receive GEM's learned representations through LoopFM and a precomputation system called Solaris, avoiding a full foundation-model pass during every live ad request. For GEM overall, Meta reports 18.9% better conversions and 12% higher click-through rates across measured surfaces and objectives. Those figures come from Meta's own evaluations; the post does not provide an independent replication or a universal advertiser outcome. The technical change matters because it puts a large shared model behind the production ad system, while smaller serving models handle latency-sensitive requests. Read Meta's dated GEM engineering account.

🔓 OPEN SOURCE

NVIDIA opens full Run:ai platform and four GPU orchestration modules

NVIDIA released the full Run:ai platform as open source for AI workload and GPU orchestration. Teams can use the integrated Kubernetes platform or adopt four components separately: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. The scheduler places GPU jobs with quota, topology and sharing controls; the fractioning operator exposes CUDA-native GPU memory fractions to Kubernetes. NVIDIA says the Community Product also ships through NGC with a Helm chart, while the code is available in its DSX GitHub organization. The release gives infrastructure teams a concrete way to inspect and modify tooling previously distributed as a commercial platform. Licensing still needs care: the source components have their applicable open-source licenses, but downloading the NGC package requires NVIDIA’s software and product terms. The Community Product includes neither enterprise support nor a paid subscription entitlement. See NVIDIA’s Run:ai release and adoption details.

💸 DEALS & FUNDING

Clueso raises $10M for agents that create and update business videos

Clueso announced a $10 million Series A on October 13 to expand its AI platform for creating and maintaining business videos. Dharana Capital led the round; Y Combinator and f7 Ventures returned, while Founders Future and Neotribe joined. The company says more than 1,000 organizations use its tools for marketing, training, and customer education. Clueso's pitch goes beyond generating a clip from a prompt. The company describes computer agents that capture product workflows and revisit them when interfaces change. Its platform design includes a video editing agent to coordinate scripts, voiceovers, motion graphics, and edits, while reusable Clueprints encode brand and localization rules. An MCP connection lets other AI assistants invoke Clueso's video tools. Clueso’s funding announcement says the capital will support agent infrastructure, the editing platform, and hiring. Those are company claims about a growing product; the announcement does not independently establish how reliably automated video updates work across customers.

Mistlabs launches Ambi after $26M funding round

Mistlabs launched Ambi on Apple devices after announcing $26 million in Pre-A funding. Luminous Ventures and Huaqin led the financing, which the company says will help fund a standalone wearable. Ambi is meant to use conversation and work context across iPhone, Apple Watch and Mac to suggest next steps as a meeting unfolds. Mistlabs describes Routines that can research a topic and prepare drafts, but has not published independent evidence that these tasks run reliably. Its proposed communication between users’ agents is scheduled for the coming months, not part of the current launch. The company says audio is not stored by default. Its privacy notice says Listening audio is deleted when listening ends, while explicitly saved recordings remain until deletion. Remaining information is encrypted and stored in the United States; processing is not entirely local. Consent for nearby people remains a practical issue with ambient capture. Read Mistlabs’ dated Ambi launch announcement.

KEY TAKEAWAYS

Google opens Gemini 4 Argon - Paid API access is live, with introductory pricing through November 30.

Meta launches Muse Wonder - Wonder is available through the Meta Model API and Muse Code; Full targets Muse Code subscribers.

Liquid expands local agent options - The 24B open-weight model activates 2B parameters per token, with company-reported speed and mixed benchmark results.

Sentient OS announces version 2.0 - A local context model feeds computer-use tasks; hosted model choices may transmit task context.

Microsoft announces Maia 300 - The new inference system has a 2027 production target; Maia 200 already serves production workloads.

Meta rebuilds GEM for ads - Its 2026 model unifies recommendation across surfaces; reported conversion and click gains are company measurements.

Run:ai code is open - NVIDIA released the full platform and modular components, with separate NGC terms and no enterprise support for the Community Product.

Clueso raises a $10M Series A - Dharana Capital led the round to fund video agents, editing infrastructure and hiring.

Mistlabs funds ambient AI - Its $26 million Pre-A backs an Apple-device launch and planned wearable, while promised agent features remain unverified.

BotsFired · Bots Keep Firing 🔥