English translation is being prepared.
Matched fieldsSummary
←Search results
Search “Search”
English translation is being prepared.
Matched fieldsSummary
English translation is being prepared.
Matched fieldsSummary
English translation is being prepared.
Matched fieldsSummary
English translation is being prepared.
Matched fieldsSummary
English translation is being prepared.
Matched fieldsSummary
Thinking Machines Lab Releases Inkling 2 Small, a 594B Total Parameter MoE Model
Thinking Machines Lab has released its second-generation model Inkling 2 Small, which uses a 594B total parameter, 17B active parameter MoE architecture, supports image input and variable reasoning effort, and scores 31 on the Artificial Analysis Intelligence Index. The model is first offered on Tinker with closed weights, and is planned to be released with full weights under the Apache 2.0 license next week after external safety testing and red-team evaluation are completed. The company says it delivers comprehensive improvements over the previous generation in long-horizon agentic coding, vision, and knowledge work, and its training includes 8.4 trillion token pretraining and joint RL with 1075 steps and over 17 million rollouts.
Thinking Machines Lab has released its second-generation model Inkling 2 Small, which uses a 594B total parameter, 17B active parameter MoE architecture, supports image input and variable reasoning effort, and scores 31 on the Artificial Analysis Intelligence Index. The model is first offered on Tinker with closed weights, and is planned to be released with full weights under the Apache 2.0 license next week after external safety testing and red-team evaluation are completed. The company says it
Matched fieldsBody
NVIDIA Releases Vera CPU, Focused on Agent Throughput and Energy Efficiency
NVIDIA has released the Vera CPU for AI agent workloads, claiming that in a 100-megawatt data center comparison, a single Vera outperforms the fastest x86 CPU and the most energy-efficient x86 CPU. NVIDIA's figures are: compared with the fastest x86 CPU, Vera can support 1.6 times the number of agents per watt; compared with the most energy-efficient x86 CPU, agent speed is 2 times faster. Early adopters including Perplexity, DeepInfra, ClickHouse, DuckLabs, Daytona, and Modal provided production environment test results, among which Daytona used Vera to launch 1,000 agent sandboxes in 10.8 seconds, versus 88.4 seconds on x86.
NVIDIA has released the Vera CPU for AI agent workloads, claiming that in a 100-megawatt data center comparison, a single Vera outperforms the fastest x86 CPU and the most energy-efficient x86 CPU. NVIDIA's figures are: compared with the fastest x86 CPU, Vera can support 1.6 times the number of agents per watt; compared with the most energy-efficient x86 CPU, agent speed is 2 times faster. Early adopters including Perplexity, DeepInfra, ClickHouse, DuckLabs, Daytona, and Modal provided productio
Matched fieldsBody
Microsoft Research uses PDD Kit to formally verify a full-stack web application written by an agent
Microsoft Research, together with GitHub Next and the RiSE team, proposes the Proof-Driven Development (PDD) paradigm and releases the research prototype PDD Kit, which turns user intent into Verus formal specifications and then proves that the agent-generated code satisfies those specifications for every input, falling back to testing for the parts that proofs do not cover. The team used PDD Kit in GitHub Next's Chopin environment to verify the login logic of the Rust full-stack application Pet Store; of roughly 5,000 lines of Rust code, developers reviewed only the intent documents and specifications and did not read the generated code line by line. PDD Kit currently supports only Rust, and 6 places in the login logic are listed as trusted boundaries, covered by testing rather than proofs.
Microsoft Research, together with GitHub Next and the RiSE team, proposes the Proof-Driven Development (PDD) paradigm and releases the research prototype PDD Kit, which turns user intent into Verus formal specifications and then proves that the agent-generated code satisfies those specifications for every input, falling back to testing for the parts that proofs do not cover. The team used PDD Kit in GitHub Next's Chopin environment to verify the login logic of the Rust full-stack application Pet
Matched fieldsBody
OpenAI releases the GPT-6 Cyber system card, a specialized model for authorized security work
OpenAI releases the GPT-6 Cyber system card, a cybersecurity-specialized variant of GPT-6 Astra intended for authorized security work such as vulnerability discovery, exploit development, and web penetration testing. In defensive-use evaluations, the model completed 100% of general security assistance and vulnerability remediation tasks, 96% of exploit development tasks, and 97% of red-team tasks. Under the Preparedness Framework, its capability ratings are the same as GPT-6 Astra: cybersecurity is Critical, biochem is High, and AI self-improvement is below High; access is restricted through Daybreak Red and requires identity verification, SSO, MFA, and an API IP allowlist.
OpenAI releases the GPT-6 Cyber system card, a cybersecurity-specialized variant of GPT-6 Astra intended for authorized security work such as vulnerability discovery, exploit development, and web penetration testing. In defensive-use evaluations, the model completed 100% of general security assistance and vulnerability remediation tasks, 96% of exploit development tasks, and 97% of red-team tasks. Under the Preparedness Framework, its capability ratings are the same as GPT-6 Astra: cybersecurity
Matched fieldsBody
Apple Proposes SupportNet and KeyNet: Using Small Neural Networks to Accelerate MIPS Retrieval and Improve RAG Efficiency
In the ICML 2026 paper "Amortizing Maximum Inner Product Search with Learned Support Functions," the Apple research team proposes training a small neural network offline for a fixed key database, treating MIPS as a support function in convex analysis to learn it, thereby routing queries faster and more accurately at inference time. The methods include SupportNet (using an ICNN to learn the support function, which can be computed per cluster to improve routing) and KeyNet (directly outputting the best-matching key vector, constrained by a consistency loss based on the Euler homogeneous function theorem). Both are trained on samples from the relevant query distribution, trading a one-time amortized cost for more efficient RAG retrieval during inference.
In the ICML 2026 paper "Amortizing Maximum Inner Product Search with Learned Support Functions," the Apple research team proposes training a small neural network offline for a fixed key database, treating MIPS as a support function in convex analysis to learn it, thereby routing queries faster and more accurately at inference time. The methods include SupportNet (using an ICNN to learn the support function, which can be computed per cluster to improve routing) and KeyNet (directly outputting the
Matched fieldsBody
Reflection releases Beam open-source model safety training and evaluation report
Reflection has released the safety technical report for Beam, describing how it balances refusing harmful requests with answering normal ones through pre-training data filtering, a safety-aligned teacher model, and multi-teacher online policy distillation (MOPD). In its self-built Iris Safety Evaluations, Beam's refusal rate for harmful requests is 89%, and its compliance rate for safe, harmless requests is 82%, placing it on the Pareto frontier of safety and helpfulness. The report also shows that after the safety-aligned teacher increased from 100B to 501B parameters, over-refusal decreased relatively by 48.7%, while the pass rate for harmful requests decreased by 3.1%.
Reflection has released the safety technical report for Beam, describing how it balances refusing harmful requests with answering normal ones through pre-training data filtering, a safety-aligned teacher model, and multi-teacher online policy distillation (MOPD). In its self-built Iris Safety Evaluations, Beam's refusal rate for harmful requests is 89%, and its compliance rate for safe, harmless requests is 82%, placing it on the Pareto frontier of safety and helpfulness. The report also shows t
Matched fieldsBody
Databricks on why enterprises are turning to open-weight models: Unity AI Gateway provides unified hosting
Databricks believes open-weight models have approached frontier-model performance at lower cost, prompting enterprises to reassess single frontier-model vendor strategies. Its Unity AI Gateway hosts open-weight, frontier, and external models through a unified API, supporting intelligent routing, fallback, and budget controls; since January, the average number of models used by customers has risen from 4.2 to 7.1, of which about 40% are open-weight models.
Matched fieldsBody
Agile Deliberation: How Google Research Uses AI to Help Humans Clarify the Gray Areas of Subjective Classification
Google Research proposes the Agile Deliberation framework, which turns the construction of subjective image classifiers into an iterative dialogue between humans and LLMs: users first break down subcategories and exclusions, then the VLM proactively selects semantically borderline images, and the feedback enters an automatic prompt optimization loop. The team validated this in 18 real-human experiments of 1.5 hours each; the process reduced user fatigue and anxiety, and the resulting classifiers outperformed standard baselines. The code and algorithms are open-sourced on GitHub.
Google Research proposes the Agile Deliberation framework, which turns the construction of subjective image classifiers into an iterative dialogue between humans and LLMs: users first break down subcategories and exclusions, then the VLM proactively selects semantically borderline images, and the feedback enters an automatic prompt optimization loop. The team validated this in 18 real-human experiments of 1.5 hours each; the process reduced user fatigue and anxiety, and the resulting classifiers
Matched fieldsBody
Smartsheet launches AI planning and execution platform vsActual, using Execution Graph to track project deviations
At the Engage 2026 conference, Smartsheet announced vsActual, an AI-driven planning and execution platform that uses AI agents to continuously monitor work progress across enterprise systems, compare it against strategic goals, and recommend adjustments. Its core Execution Graph is built on roughly 20 years of workflow data and 9 PB of information, underpinned by an AWS Neptune knowledge graph of about 100 million nodes, and can connect to systems such as Oracle, SAP, Workday, Atlassian, and ServiceNow. The platform also adds Agentic Workflows, multiple integrations, and ISO/IEC 42001 certification, while retaining human approval by default.
At the Engage 2026 conference, Smartsheet announced vsActual, an AI-driven planning and execution platform that uses AI agents to continuously monitor work progress across enterprise systems, compare it against strategic goals, and recommend adjustments. Its core Execution Graph is built on roughly 20 years of workflow data and 9 PB of information, underpinned by an AWS Neptune knowledge graph of about 100 million nodes, and can connect to systems such as Oracle, SAP, Workday, Atlassian, and Ser
Matched fieldsBody
How Figma Manages Sandboxes at Scale: An Analysis of the Nimbus Platform
Figma built a centralized sandbox management platform called Nimbus. Through namespace configuration files and three APIs—GetOrCreate, Get, and Delete—engineers can create, reuse, and delete sandboxes without needing to understand cloud provider details. Nimbus uses a unified base image, sandbox-base, and a daemon, sandboxd, to manage processes, networking, and health checks. Its goal is to return a fully configured sandbox within 1 second and to load-balance across multiple cloud providers to handle failures.
Figma built a centralized sandbox management platform called Nimbus. Through namespace configuration files and three APIs—GetOrCreate, Get, and Delete—engineers can create, reuse, and delete sandboxes without needing to understand cloud provider details. Nimbus uses a unified base image, sandbox-base, and a daemon, sandboxd, to manage processes, networking, and health checks. Its goal is to return a fully configured sandbox within 1 second and to load-balance across multiple cloud providers to h
Matched fieldsBody
How Indegene Achieved Enterprise-Scale Data Quality Modernization in Life Sciences with an LLM Data Quality Agent
Indegene built an LLM-powered Data Quality Agent within the NEXT Unified Data Platform (NEXT UDP) on AWS, which can generate context-aware data quality rules, perform validation, and produce directly runnable SQL and Python code, all under centralized human-in-the-loop governance. NEXT UDP also provides independent AI Agents such as Data Ingestion, Orchestration, Data Catalog, and SQL Bot Generator, which can be used separately in data pipelines or customized for agent-to-agent communication.
Matched fieldsBody
Claude Marketplace adds security, productivity, and creative agents and products
Claude Marketplace adds partner agents and products spanning security, development infrastructure, creativity, and productivity, bringing the total number of partner agents and products on the platform to more than 25. Eligible enterprise customers can apply a portion of their Anthropic committed spend to these solutions, sharing a single committed amount across Claude and partner products. New products include Chainguard Harness, Checkmarx One, Datadog Agent Observability, Runway, Asana AI Teammates, and Zoom Agentic Search, among others.
Claude Marketplace adds partner agents and products spanning security, development infrastructure, creativity, and productivity, bringing the total number of partner agents and products on the platform to more than 25. Eligible enterprise customers can apply a portion of their Anthropic committed spend to these solutions, sharing a single committed amount across Claude and partner products. New products include Chainguard Harness, Checkmarx One, Datadog Agent Observability, Runway, Asana AI Team
Matched fieldsBody
Cohere Retrieval Models Embed 5 and Rerank 5 Now Available on OpenRouter
Cohere's Embed 5 and Rerank 5 retrieval models are now available on OpenRouter, accessible through a single API call with unified billing and no additional SDK required. Embed 5 Pro and Fast share the same embedding space, so you can index with Pro and query with Fast to reduce latency; Rerank 5's scores are calibrated, allowing the same relevance threshold to filter weak context across queries and domains. Embed 5 supports 256 to 2048-dimensional Matryoshka vectors, defaults to 2048 dimensions, and handles up to 128k tokens per request.
Cohere's Embed 5 and Rerank 5 retrieval models are now available on OpenRouter, accessible through a single API call with unified billing and no additional SDK required. Embed 5 Pro and Fast share the same embedding space, so you can index with Pro and query with Fast to reduce latency; Rerank 5's scores are calibrated, allowing the same relevance threshold to filter weak context across queries and domains. Embed 5 supports 256 to 2048-dimensional Matryoshka vectors, defaults to 2048 dimensions,
Matched fieldsBody
Together AI Releases Together Compiler, Using Agents for Continuous Compilation and Inference Optimization
Together AI has released Together Compiler (TKCC), a compilation system that uses high-throughput agents to continuously optimize inference workloads, achieving up to 2.6x TPS improvement within less than 24 hours on frontier open-source models such as Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash. On SemiAnalysis's InferenceX low-latency benchmark, it achieved current best results of 1111 TPS for MiniMax M3.1-Flash and 1205 TPS for DeepSeek v4.1-Flash, and in the Llama 1B batch size 1 scenario it is 1.7x higher than the open-source megakernel. TKCC relies on a persistent engine to hot-load candidate forward computations and a fast/slow two-tier engine to balance throughput and fidelity; its experimental throughput is 252x higher than baseline agent optimization systems, and it is currently used to serve trillions of tokens per day.
Together AI has released Together Compiler (TKCC), a compilation system that uses high-throughput agents to continuously optimize inference workloads, achieving up to 2.6x TPS improvement within less than 24 hours on frontier open-source models such as Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash. On SemiAnalysis's InferenceX low-latency benchmark, it achieved current best results of 1111 TPS for MiniMax M3.1-Flash and 1205 TPS for DeepSeek v4.1-Flash, and in the Llama 1B batch
Matched fieldsBody