AI progress - 26 october 2026. Kimi K3.1 Pro is the main release of the weekend: it is live on Moonshot's API, with open weights promised by 4 November. Until the weights ship, it is an API-only model. Elsewhere, Ollama added speech output on macOS, and Apollo published a sceptical note on AI returns outside the tech sector. Releases were otherwise light. New models and papers. * Kimi K3.1 Pro and K3.1 Pro Highspeed - Moonshot's new model is on the Kimi API, with a faster "Highspeed" variant. Why it matters: weights are not open yet, but Moonshot promises a release by 4 November 2026, so self-hosting is a date to plan around, not an option today. * RynnUnify: Towards Unified Embodied Brain - a 16B model that combines embodied understanding with long-horizon visual planning through interleaved video-language reasoning. Why it matters: it is one model for both perception and planning in robotics, where those are often separate; the results are the authors' own and unreplicated. * HoloBoxing: A Simple Recipe for Humanoid Boxing - the authors build on a pretrained whole-body controller and separate reusable motor skills from the task, reporting 7 minutes of single-GPU training. Why it matters: if the claim holds, new humanoid tasks can be added on top of an existing controller at very low training cost. New tools and software. * Ollama v0.40.11 - adds speech generation on macOS; ollama run plays text or a chat model's reply aloud as audio is generated. Why it matters: local models can now talk back without a separate TTS stack, at least on Mac. * Claude Code v2.1.310 - the Monitor tool and plugin monitors now work on Bedrock, Vertex and Foundry, and when nonessential traffic is disabled. Why it matters: teams on those cloud providers or locked-down networks get the same monitoring workflow as direct users. * Copilot usage metrics API adds tools like OpenCode - the API now reports model usage from more tools developers use with Copilot. Why it matters: admins measuring Copilot usage get a fuller count than before. Overall digest. * No signs of AI payoff outside the tech sector - an Apollo note arguing that most companies outside tech see no financial return from AI; TechCrunch reports the economist's phrase as "zero financial payoff". Why it matters: this is one firm's view, but it is a concrete claim to test against your own deployment's numbers. * Pennsylvania will require labels on AI-generated ads - Governor Shapiro signed the Artificial Intelligence Deceptive Advertising Act, requiring disclosures on AI-generated ads. Why it matters: anyone producing ad content for Pennsylvania audiences now has a disclosure obligation to build into their workflow. * Securing its frontier model training environment - OpenAI describes stronger containment for training and evaluation, including offline environments and high-compute red-teaming. Why it matters: it is a primary-source look at how a frontier lab isolates training runs, useful as a reference for your own controls. Get the AI digest by email. One email per weekday morning: new models and papers, new tools, and the overall picture. No ads, unsubscribe in one click. See the privacy policy.

Reflection Beam on Ollama: 501b-parameter open-weight MoE model, free for a limited time. October 16, 2026 · Originally published October 15, 2026 · deepseek/deepseek-v4.1-flash Reflection released Beam, its first open-weight model, on October 16, 2026, and Ollama shipped it as a day-zero partner on Ollama Cloud - free for a limited time. The release matters because Beam is an Apache 2.0 mixture-of-experts model with roughly 501 billion total parameters, 23 billion active parameters per token, and a 1 million token context window, aimed squarely at software engineering and long-horizon agentic workloads rather than single-turn chat. What was announced. Beam is Reflection's first open-weight model and became generally available on the morning of October 16, 2026. Ollama is a day-zero launch partner, making Beam available through Ollama Cloud at no cost for a limited promotional period. Ollama's Beam endpoint runs with full zero data retention (ZDR) in collaboration with Reflection. Ollama Cloud lets developers run supported models from their apps or terminal without downloading models or managing local inference. Model architecture and specifications. Beam is a mixture-of-experts model with approximately 501 billion total parameters and 23 billion active parameters per token. The gap between total and active parameters is the defining efficiency characteristic of a sparse MoE design: only a fraction of the network is engaged per token. | Property | Value | | Total parameters | ~501 billion | | Active parameters per token | 23 billion | | Architecture | Mixture-of-experts | | Context window | 1 million tokens | | Input/output modality | Text-in, text-out | | License | Apache 2.0 | Beam's 1 million token context window is explicitly positioned to support long-running workflows. Its text-in, text-out design focuses on workloads where code, tool calls, logs, documents, and structured outputs are central. What Beam is designed to do. Beam targets workflows that involve more than a single response - writing and maintaining code, using tools across multiple steps, and working through complex tasks over time. Ollama lists four primary workload categories: * Coding agents: solving software engineering tasks, working through codebases, and helping write and maintain code. * Long-running agents: executing multi-step workflows, using tools, and working toward a goal over an extended sequence of actions. * Math and science reasoning: working through complex mathematical and scientific problems. * Custom models: fine-tuning and customizing an open-weight model for specific applications. Benchmarks. Ollama's announcement includes four benchmark charts covering agentic coding, reasoning, tool calling, and general performance. The blog post publishes these as images and does not state numeric scores in text, so no specific benchmark figures are available from this source. How to run beam. Beam is accessed through Ollama Cloud using the beam:cloud tag. Ollama directs users to download the latest version of Ollama first. Run Beam directly: ollama run beam:cloud Launch it inside coding agents: ollama launch claude -model beam:cloud ollama launch codex -model beam:cloud ollama launch opencode -model beam:cloud Why the Apache 2.0 license matters. Beam is released under Apache 2.0, which gives developers the freedom to inspect, customize, and build on the model. Combined with the open-weight release, this permits fine-tuning for specific applications - a use case Ollama explicitly lists among Beam's intended workloads. Key takeaways. * Reflection's Beam is the company's first open-weight model, generally available as of October 16, 2026. * Beam is a mixture-of-experts model with ~501B total parameters and 23B active parameters per token. * It offers a 1 million token context window and is text-in, text-out. * It is released under Apache 2.0. * Ollama is a day-zero partner offering Beam free for a limited time via Ollama Cloud, with full zero data retention. * The model is positioned for coding agents, long-running agents, math and science reasoning, and fine-tuning into custom models. Further information.

Ollama v0.40.3 removes warnings for Claude and Codex launches. 2026-10-11 Ollama has released version 0.40.3 of its local AI tool, addressing user experience issues with model launches. The update removes warning messages that previously appeared when launching Claude and Codex models through the ollama launch command. These warnings were considered unnecessary friction for developers who already understand the context of the models they're running. The changes are documented in the official release notes at github.com/ollama/ollama/releases/tag/v0.40.3. The primary change in this release eliminates the warning that appeared when using ollama launch claude, which was previously flagged due to Claude connector configurations. Similarly, ollama launch codex no longer displays warnings about service tiers. These adjustments streamline the command-line interface for developers who frequently switch between different AI models for various tasks. The modifications come from pull request #18878, addressing issue #18878 in the Ollama repository. For developers building applications that require multiple AI models, these changes represent a meaningful improvement to daily workflows. When working with local AI infrastructure, every warning message that appears during model initialization adds cognitive load and potential confusion. The removal of these warnings suggests that Ollama is refining its interface to better serve users who need reliable, predictable behavior when scripting automated processes or building development tools that depend on local model execution. The changes also indicate that Ollama continues to iterate based on user feedback. Rather than maintaining strict warnings as a default, the team has adjusted the behavior to align with what experienced users would expect. This approach recognizes that developers using Ollama often have specific knowledge about their model configurations and service requirements, making generic warnings more of an obstacle than a help. Version 0.40.3 does not introduce new model support or major architectural changes. Instead, it focuses on refining the existing user experience around model launching. Users who have encountered these warnings in their scripts or automated workflows will notice cleaner output and fewer interruptions. The update is available through standard Ollama installation methods, and existing installations can upgrade without requiring reconfiguration of their model setups. This release reflects a maturation of Ollama's approach to user interface design. Early versions of local AI tools often included extensive warnings and guidance, assuming users might not understand the implications of their choices. As the ecosystem has grown and users have become more sophisticated, the interface has evolved to reduce unnecessary friction while preserving important safeguards where they're genuinely needed. Developers building applications that integrate with Ollama should test their workflows with version 0.40.3 to ensure compatibility with the updated launch behavior. The changes are unlikely to break existing functionality, but automated systems that parse command output might need adjustment if they were previously handling the removed warnings. Ollama remains a self-contained tool for running language models locally, separate from any Mina Labs offerings. The improvements in this release enhance the core experience of working with local AI models without changing the fundamental capabilities of the platform. For teams developing applications that rely on local model execution, these refinements contribute to a more polished and predictable development environment.