Cainew

A 9-billion parameter open-source model was fine-tuned with reinforcement learning for $500 and achieved performance comparable to frontier AI models on catalog review tasks. This demonstrates the cost-effectiveness of fine-tuning smaller models for specialized applications.

Neutrino-1 8B is a new AI model offering 8 billion parameters with optimized performance characteristics. The model represents advances in efficient AI model design and deployment.

Claude's Opus 5 model is benchmarked against SlopCodeBench, a coding evaluation suite. The results demonstrate Opus 5's performance capabilities in code generation and solving programming tasks.

Most agents pass their evals and fail in production. Prefactor is the evaluation layer that closes the gap. We score every agent run in real time, surface quality regressions and drift as they happen, and show engineering teams exactly how their agents are performing at scale. Built for the teams shipping agents to customers.

Kimi K3, an advanced AI model, is now accessible through the Telnyx Inference API for developers and businesses. This integration enables easier deployment and usage of Kimi K3 capabilities via Telnyx's infrastructure.

We allow companies that operate in the physical world (home remodeling, roofing, trades) to automate inbound and outbound calling & texting and run multi-threaded campaigns spanning several weeks. Our AI agents can hold 100+ parallel conversations, speak in multiple languages and integrate into any CRM. We help put appointment scheduling, customer service & confirmation calls on autopilot and eliminate missed calls, slow speed to lead and insufficient lead follow ups for our customers.

Company Brain for AI agents, Fragmented communication and data of a company which is left on channels Emails, Slacks, WhatsApp and LinkedIn connect it all in one place with approval layer attach so personal and works chats don't gets mixed and connect it with any ai agents simultaneously, where records are keep getting updated minutes by minutes and so AI agents are getting updated as well so use AI agents with less cost of context and don't have to repeat.

Every time someone gets hired, applies for a mortgage, or leases an apartment, someone has to confirm they actually worked where they said they did. That means millions of calls, emails, faxes (yes, in 2026), to HR and payroll departments. Superunit's AI agents do all of that from start to finish. They research, call, email, fax and handle documents until the verification is done.

Conduit automates guest conversations and the work behind it. Conduit agents reply across email, WhatsApp, OTAs and social, voice calls. Agents dispatch subagents to act inside your PMS and tech stack: dispatching cleaners, updating calendars, closing the loop. Agents learn from every interaction, and you can inspect every step. 140+ languages, SOC 2 Type II. 50M+ conversations and $3B+ in reservation value for 300+ brands including Marriott, Hilton, Nobu and Fairmont.

Your agent writes files locally in markdown and HTML, the formats it already speaks. Liminal renders them as clean UI instantly and syncs to your team in real time. Share a link and everyone sees the same live workspace, editable in the browser. No sending markdown over Slack, no hosting your own HTML, no waiting on Google Drive. No MCP calls, no proprietary formats, no wasted tokens. Over time, it becomes a shared second brain for your team and every agent working with you.

Phantom is a voice-first AI agent for macOS that is always accessible via the Mac notch. Instead of opening a chatbot window and moving work into a separate conversation, users speak or type from anywhere on their Mac, without interrupting their work. Phantom understands the relevant screen and app context and completes tasks directly within the user's workflow. The possibilities are only limited by the imagination of the user.

Recipe Book is a video data platform where you can semantically search 25M+ clips, iterate on the results, and buy the exact dataset you need to train your model in one sitting. Search -> rate a few clips -> we train a probe on your votes to re-rank the whole catalog to your taste. No contact forms, no sales calls, $3/hour.

We’re announcing even more new capabilities in Managed Agents in Gemini API so developers can build reliable, production-ready agents.

Google introduces Beyond Zero, an enterprise security framework designed for the AI era. The initiative addresses security challenges specific to AI deployment in enterprise environments.

XY is a new GPU-accelerated interactive plotting library designed for fast, composable visualizations. It enables efficient data visualization with support for GPU computation.

This article argues that LLMs should be granted access to the ACM Digital Library to improve their training and capabilities. Providing access to quality academic resources would enhance AI research and development.

It might sound counterintuitive, but Search's AI tools can actually help you make the most of your time offline whether you want to book concert tickets or find the perf…

PyTorch is discussed as a reference language for AI development and implementation. The article explores PyTorch's role as a foundational framework in the AI and machine learning ecosystem.

Multimodal foundation models learned to see and to speak by consuming the whole internet. Embodied agents admit no such shortcut, since they require data that couple observations with physical states and actions. These signals can be provided, to varying degrees, by multiple data sources. In this work, we organize the embodied data ecosystem as a "pyramid" spanning five complementary sources: real-robot data, UMI-style data, egocentric and exocentric data, simulation data, and general vision-lan...

On-policy distillation (OPD) adapts diffusion models by querying a teacher along trajectories generated by the current student, but how it should behave under classifier-free guidance (CFG), a default component of modern diffusion systems, remains poorly understood. Existing OPD methods naturally extend velocity matching to the CFG-composed prediction, directly matching teacher and student guided velocities. We show that this objective is under-identified at the branch level: positive- and negat...

Reliable visual document understanding requires a model to attribute each answer to the evidence regions that support it. Recent benchmarks and systems express this step through a coordinate interface: the model outputs the coordinates of bounding boxes that mark the evidence regions in the document. Under this interface, vision-language models often fail to identify the right regions even when the answer is correct, a failure known as Attribution Hallucination. We present a study that investiga...

Progress in video generation keeps narrowing the visual gap between AI-generated and professionally produced footage, yet most benchmarks still draw prompts from web sources or LLM templates and score them with untrained, generic multimodal models. More fundamentally, their evaluation taxonomies remain rudimentary (overall visual quality, coarse text alignment and temporal smoothness) rather than the professional Cinematic Language criteria by which films are actually made and judged, so they as...

Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced proprietary models with their strong reasoning capabilities are promising teachers. While distilling from proprietary models can densify this supervisory signal, conventional logit-matching is precluded...

Many recent robot policies pursue stronger control by using large pretrained vision-language models (VLMs) as the action backbone. We introduce WorldDiT, a unified diffusion transformer architecture that couples action generation with visual world modeling and achieves strong performance without a large pretrained VLM action backbone. During training, a single diffusion transformer generates continuous action chunks and predicts normalized RGB patch targets from future camera frames. Across four...

Multimodal large language models (MLLMs) hold immense potential to revolutionize clinical practice, yet deploying them in the medical domain is fundamentally a vision-centric challenge: models must absorb knowledge from heterogeneous 2D and 3D medical images, and evaluation protocols must align with radiologists' clinical practice and provide an accurate, fine-grained and factualness-driven assessment. In this paper, we introduce ClinFusion, a vision-centric MLLM designed for holistic medical un...

While data curation for Vision Language Models (VLMs) is increasingly active, public practice for constructing pretraining mixtures remains largely heuristic: practitioners stack datasets that pass quality filters, set cross-domain ratios by intuition, and lack a principled, attributable criterion for admitting new data, while frontier recipes remain undisclosed. We formulate data construction as a systematic mixture-optimization problem and turn it into a reproducible engineering discipline by ...

Multi-turn long-horizon planning is critical for foundation model agents, yet how to fundamentally improve it remains unclear. Existing models are trained on uncontrollable and opaque Internet data, making it difficult to identify how planning ability is acquired, shaped, and integrated. To address this challenge, we introduce a unified and controlled multi-turn environment that enables precise control. It allows systematically study long-horizon planning across three stages. (1) Planning abilit...

Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected key-value blocks, yet existing methods struggle to sparsify attention both efficiently and accurately for two reasons: (1) Rigid, unpredictable, and costly routing: selecting a fixed fraction of top-ranked blocks by proxy score imposes fixed budgets, whereas re...

A new field report shows how scientists use AI coding agents to modernize scientific computing, accelerating software development and discovery in genomics and beyond.

JFrog and OpenAI's research on zero-day vulnerabilities emphasizes that rapid remediation has become more important than prevention for maintaining trust. Fast response times to security issues are now the key metric for trustworthiness.

This article presents a position statement on open-weights models and their role in the AI ecosystem. It discusses the benefits and considerations of releasing model weights openly to the community.

Researchers demonstrate how Claude can be used to discover cryptographic weaknesses and vulnerabilities in security systems. The findings highlight AI's potential in identifying security flaws through analysis.

Large language models should not be relied upon to provide confidence scores for their outputs. The article argues against using LLM-generated confidence metrics due to their unreliability and potential to mislead users.

A developer shares their approach of not reviewing code generated by AI agents to understand trade-offs in agent-driven development. The strategy highlights a hands-off approach to agentic AI usage despite potential risks and benefits.