1. Google Launches Gemini Distillation Service
Google has launched the Gemini Distillation Service, a managed offering designed to transfer the reasoning patterns of its frontier models into highly efficient student models. The service currently supports distilling knowledge from gemini-3.1-pro into gemini-2.5-flash. This allows developers to build specialized, low-latency models that retain advanced reasoning capabilities while significantly reducing inference costs for production applications.
- • The Gemini Distillation Service automates the training of smaller student models using larger teacher models.
- • The service currently supports gemini-3.1-pro as the teacher model and gemini-2.5-flash as the student model.
- • The service is optimized for high-volume, latency-sensitive applications that require complex reasoning capabilities.
Developers can now easily distill complex reasoning capabilities into smaller, faster, and cheaper models for high-volume, latency-sensitive applications.
2. Microsoft Releases Mage-VL Streaming Multimodal Model
Microsoft has released Mage-VL, an open-weights codec-native streaming multimodal foundation model optimized for image and video understanding. Mage-VL pairs a custom 4B-scale visual encoder with a Qwen3-4B language backbone. By separating video streams into anchor and predicted frames, the model's sparsity mechanism filters out redundant visual tokens by over 75%, yielding a 3.5× speedup in inference time. Additionally, a built-in 'cognition gate' ensures the full vision-language model is only invoked when a significant event is detected, drastically reducing unnecessary compute.
- • Mage-VL is a codec-native, proactive-streaming multimodal foundation model for image and video understanding.
- • The model features a 4B-scale visual encoder trained from scratch and a Qwen3-4B-Instruct-2507 language backbone.
- • A codec-aligned sparsity mechanism separates video streams into anchor (I) and predicted (P) frames, reducing visual tokens by over 75%.
- • The architecture achieves up to a 3.5× wall-clock inference speedup compared to uniform frame sampling.
- • A dual-process design uses a 'cognition gate' to invoke the full VLM only when a response-worthy event is detected.
Developers can build highly efficient video-processing pipelines that drastically reduce inference latency and token consumption.
3. Runway Details Characters Real-Time Video Model and Distillation Workflows
At VB Transform 2026, Runway shared production insights from its real-time video model, Runway Characters, which enables zero-latency avatar interactions. To achieve real-time speeds, Runway distills large foundation models into smaller student models, cutting generation latency by 80% to 90% while using adversarial post-training (APT) to preserve visual sharpness. The team also detailed practical engineering workarounds, such as using a frontend re-centering feature to resolve character drift bugs and deploying AI monitoring agents to isolate hardware-level performance drops in cloud data centers.
- • Runway Characters is a real-time video model enabling zero-latency interactions with AI-generated avatars.
- • Runway utilizes distillation to train smaller student models, reducing video generation times by 80% to 90%.
- • The team employs adversarial post-training (APT) to maintain visual sharpness in distilled models.
- • Runway resolved a character drift bug by implementing an 'Optimize for Image Quality' frontend feature that re-centers input images.
- • The company used AI agents and monitoring tools to trace an API slowdown (8% of calls dropping to 16 fps) to faulty hardware in us-east-1.
Developers building interactive video applications can learn from Runway's production strategies to optimize latency and handle character drift.
4. Microsoft Expands MAI Model Family with Cybersecurity-Focused MAI-Cyber-1-Flash
Microsoft has expanded its proprietary MAI model family, first introduced at Build 2026, with the release of MAI-Cyber-1-Flash. This new 5-billion active parameter model is a specialized fine-tune of the previously released MAI-Code-1-Flash, designed specifically for defensive cybersecurity tasks like codebase patching. The model powers Microsoft's MDASH scanning harness, enabling automated security workflows while reducing operational costs.
- • MAI-Cyber-1-Flash is a cybersecurity-specialized fine-tune of the MAI-Code-1-Flash model.
- • The model features 137B total parameters, 5B active parameters, and a 256k context length.
- • It powers the MDASH scanning harness, which achieved a 95.95% score on the CyberGym benchmark.
- • The model is restricted to defensive tasks and cannot generate exploits, scoring zero on ExploitGym.
- • Integration into MDASH reduces operational costs by 50% by handling routine tasks locally before escalating to larger models.
This release marks the first cybersecurity-specific expansion of the MAI model family, providing developers with a cost-effective, defensive tool for automated security patching.
5. Model Context Protocol Officially Launches Stateless Architecture
The Agentic AI Foundation (AAIF) has officially released the final version of the Model Context Protocol (MCP) specification, transitioning from the release candidate announced in May to a fully stateless request/response architecture. This GA release enables developers to deploy MCP servers behind standard load balancers and Kubernetes, while introducing mandatory security hardening and a 12-month deprecation policy for legacy features.
- • The final MCP specification is now GA, moving from the stateless release candidate to a production-ready architecture.
- • Stateless design supports deployment behind standard load balancers and Kubernetes.
- • Includes mandatory RFC 9207 issuer validation for security.
- • Establishes a formal 12-month deprecation policy for legacy features like HTTP+SSE.
- • Graduates two new official extensions: MCP Apps and MCP Tasks.
This official release confirms the transition to a stateless architecture, allowing production-grade deployment of MCP servers in cloud-native environments.
6. Snowflake Launches Cortex AI Gateway to Govern AI Agents
Snowflake has announced the Cortex AI Gateway, a centralized governance and security layer designed to manage autonomous AI agents. Built on technology from Snowflake's acquisition of Natoma, the gateway supports over 100 Model Context Protocol (MCP) servers, allowing developers to centralize access policies and permissions. To ensure accountability, the platform implements dual attribution—logging both the agent's identity and the authorizing human—while providing finance teams with real-time cost attribution and hard spending limits to prevent runaway API bills.
- • Snowflake's Cortex AI Gateway serves as a centralized control layer for governing AI agent access to enterprise data, tools, and models.
- • The gateway supports over 100 Model Context Protocol (MCP) servers to centralize authentication, permissions, and audit logging.
- • It features dual attribution, logging both the non-human identity of the agent and the specific human who authorized the task.
- • Security integrations are included with 1Password, Aembit, Linx Security, SailPoint, and Saviynt.
- • The gateway provides unified cost attribution and spending limits to prevent runaway enterprise API costs.
Developers building enterprise agents can use this gateway to enforce security policies, manage permissions, and prevent runaway API costs.
7. Fireworks AI Launches Fireworks Nexus Routing Layer
Fireworks AI has announced Fireworks Nexus, an intelligent routing and cost-control platform designed to optimize LLM expenditures. The platform's core open-source component, FireConnect, allows developers to integrate Fireworks models into existing developer tools like Claude Code with a single-line installation. By using a custom-trained model to evaluate the difficulty of incoming requests, the intelligent router automatically offloads routine coding tasks to cost-effective open-weight models while reserving complex queries for frontier APIs, resulting in a reported 33% cost reduction per merged pull request.
- • Fireworks Nexus is an AI management and routing platform featuring enterprise cost controls and traffic management.
- • The FireConnect component is open-source under an Apache 2.0 license, enabling single-line integration with Claude Code, Codex, and OpenCode.
- • An intelligent router uses a custom model to score request difficulty, routing simple tasks to open-weight models and complex ones to frontier APIs.
- • Preliminary customer testing showed a 33% reduction in cost per merged pull request.
- • The router is in research preview and supports configurations such as routing between Claude Opus 5 and GLM-5.2.
Developers can integrate this routing layer with a single line of code to automatically reduce API costs for routine coding tasks.
8. OpenAI Makes Codex Security Tools Publicly Available via CLI and SDK
OpenAI has transitioned its Codex-powered security initiative from the previously announced 'Daybreak' closed preview to a general-purpose developer tool. The new @openai/codex-security package provides a CLI and TypeScript SDK that allows developers to automate vulnerability detection and remediation within their own build workflows. Unlike the earlier sales-gated agentic preview, this release supports non-interactive CI/CD integration using an API key, alongside interactive scanning via ChatGPT authentication.
- • Transitions Codex-based security from the 'Daybreak' closed preview to a public CLI and SDK.
- • Supports automated vulnerability scanning and remediation in CI/CD pipelines via API key.
- • Includes a programmatic CodexSecurity class for custom integration.
- • Requires Node.js 22+ and Python 3.10+.
This release moves OpenAI's security capabilities from a restricted enterprise preview to a self-service developer tool, allowing for automated security integration in standard CI/CD pipelines.
9. Visa Open-Sources Agentic Vulnerability Scanning Harness
Visa has open-sourced the Visa Vulnerability Agentic Harness, a multi-model security pipeline the company developed to hunt for bugs in its global payment infrastructure. The harness automates complex security workflows including code ingestion, threat modeling, and exploit chain synthesis. Alongside the release, Visa published a white paper outlining 12 architectural practices for critical infrastructure, emphasizing that autonomous agents must operate under distinct, scoped identities to maintain security boundaries.
- • The Visa Vulnerability Agentic Harness is now open-source on GitHub, managing code ingestion, threat modeling, and exploit synthesis.
- • The harness utilizes a multi-model pipeline, which Visa originally paired with Anthropic's Claude Mythos model to scan its global payment network.
- • Visa introduced Mean Time to Adapt (MTTA) as a metric to measure how quickly teams can confirm, fix, and validate the closure of attack paths.
- • The project's architectural guidelines mandate that AI agents be treated as distinct identities with strictly scoped permissions.
Developers can adopt this proven, multi-model pipeline to automate threat modeling and exploit chain synthesis within their own codebases.
10. Molt Agentic Reinforcement Learning Framework Released
The open-source community has introduced Molt, a PyTorch-native framework designed to treat AI agents as structured programs. Molt provides developers with a robust environment for training agents using reinforcement learning, featuring native support for custom Python rewards, tool integration, multimodal environments, and LLM-based evaluation. Built on a high-performance stack of Ray, vLLM, NVIDIA AutoModel, and FSDP2, the framework is designed to scale training up to trillion-parameter mixture-of-experts models.
- • Molt is a PyTorch-native framework that treats an AI agent as the program itself.
- • The framework supports custom Python rewards, tool use, multimodal environments, and LLM judges.
- • Molt's technical stack integrates Ray, vLLM, NVIDIA AutoModel, and FSDP2.
- • The framework is capable of scaling training up to trillion-parameter mixture-of-experts models.
Developers can use this framework to build and train complex agentic workflows with custom rewards, tool use, and LLM-in-the-loop judges.
11. Agenta Launches Open-Source Claude Cowork Alternative
Agenta has launched an open-source, self-hosted alternative to Claude Cowork, giving developers a flexible platform to build and deploy AI coworkers. Available on GitHub, Agenta allows users to run scheduled or event-triggered automations using their choice of model harness—including Claude Code, OpenAI Codex, or local self-hosted models. AI agents are configured through natural chat interactions and maintain isolated workspaces, custom tools, and self-improving configurations based on user feedback.
- • Agenta is an open-source, self-hosted alternative to Claude Cowork, available on GitHub.
- • The platform allows developers to build AI coworkers that run on schedules or trigger based on external app events.
- • It supports multiple execution harnesses, including Claude Code, OpenAI Codex, and self-hosted models.
- • Agents are created via chat, maintaining individual instructions, tools, skills, and dedicated work folders.
Developers can build and host their own AI coworkers locally, avoiding vendor lock-in while utilizing existing API subscriptions or self-hosted models.