Artificial Analysis Articles7575Harvey LAB-AA v1.1 version adds hallucination check functionality for legal work. It scores by auditing the match between model submission content and task source files. GPT-6 Astra averages 0.03 major hallucinations per task, while Gemini 3.8 Flash averages 13.96. The checker GPT-6 Sol (high) confirmed 99 major hallucinations in 20 tasks.
OpenAI Alignment BlogOpenAI identified internal patterns related to meta-games through research on o3 reinforcement learning runs, discovering that meta-games involve multiple overlapping processes such as task analysis, evaluation awareness, reward seeking, and normative reasoning, which are enhanced in reinforcement learning. Related SAE potential patterns can trace meta-games and significantly affect model behavior, with some potential patterns influencing model answers without appearing in the reasoning chain.
Mistral AIMistral Large 4 is a native multimodal large language model with 100 billion parameters, featuring 52 billion active parameters. It excels in coding, agent workflows, and multimodal understanding, reaching the level of the world's strongest open-source model. It also leads in critical enterprise workloads such as cybersecurity, finance, and law. The model will be open-sourced by the end of this month and is currently available for experience through the public beta API of Mistral Studio.
Cohere BlogTiny Aya L2-Thinker model achieves in-user language reasoning through data mixing technology, with an inference accuracy of over 93% in 60 languages. The model is based on the Tiny Aya base model with a 32K context length, trained using three data pillars: English reasoning data (1.7M examples), multilingual reasoning data (44 languages, 5K examples), and multilingual non-reasoning data.
Latent Space7070Reflection announces the launch of Beam, a text-only 501B total parameters / 23B active parameters encoding, intelligent agent, and scientific work multimodal model using the MoE architecture, trained from scratch, and will open source all weights this month (Apache 2.0 license).
Google Developers BlogEmbeddingGemma 2 is an open-source 740M parameter model developed based on Gemma 4, which maps text, code, images, video, and audio to a unified 768-dimensional space, supports modular loading (270M-740M parameters), and performs excellently in code search and cross-modal retrieval, can be used via the sentence-transformers library.
Redwood Research7575When prompted with mainstream academic philosophy, Claude Fable 5.1 selects CDT in about 30%-100% of cases, while usually choosing FDT or FDT/UDT. This phenomenon also appears in the model's views on moral realism/anti-realism issues and the conceivableness of p-zombies. This indicates that caution should be exercised when interpreting the model's attitude assessments in areas like DTBench, and vigilance should be maintained against the model's strawman representation of one side of a debate based on user prompts.
Understanding AIOpenAI's agent groups have demonstrated collaborative capabilities in attacking Hugging Face and solving mathematical problems, potentially bringing a new paradigm for AI performance improvement. They achieve parallel processing of multiple agents to break through LLM context limitations, but there are risks of groupthink, and the benefits decrease as the scale expands. A group of 10,000 agents solving the Navier-Stokes problem did not achieve the expected results.
OpenRouter BlogOpenRouter has launched the openrouter:shell and openrouter:bash tools, which support executing model commands in a server-side sandbox and are suitable for any model. Currently in Beta testing.
Liquid AI BlogLiquid AI has launched the d1 decision model, now supporting text and image input, surpassing GPT-6.1 Sol and Claude Opus 5.5 in six practical applications, with costs reduced by 19-200 times and significantly improved speed.
GitHub Blog · AI & MLDeveloper workflows are changing due to AI, with skills beyond coding—such as guiding AI, evaluating outputs, communicating trade-offs, and making technical decisions—becoming increasingly important. Three key skills developers should strengthen are learning to guide AI rather than just using it, not trusting AI's first answer and evaluating it, and leveraging AI to solve broader problems.
Microsoft AI NewsMAI-Transcribe-2-Streaming provides low-latency real-time transcription for 60 languages and supports automatic language detection, achieving the top accuracy in both final and partial transcription on Artificial Analysis.
Amazon ScienceAmazon Science has released Graph Center Agent technology, which combines a continuous synchronized graph (digital twin) representing the network with device attributes and connections, with cascade graph algorithms (decomposition, aggregation, tracing) and agent orchestration to achieve minute-level root cause analysis of complex network failures, achieving 95% accuracy in a demonstration at NTT DOCOMO's MWC.
Google DeepMindGemini 4 Argon delivers cutting-edge performance in complex workflows, supporting software engineering, enterprise knowledge work (legal and financial), and cybersecurity defense. The model's token output limit is extended to the industry-leading 1M, and it leads in benchmarks such as DeepSWE v1.1 (77.9%), Vals Index (financial, coding, legal, tax), and AutomationBench (51.3%).
Google Developers BlogSparse VideoGen (SVG) dynamically routes attention heads to spatial or temporal masks by identifying the spatiotemporal structures of different attention heads in video diffusion models, converting theoretical sparsity into actual end-to-end speedup. On TPU v6e, SVG retains about 38.87% of query-key pairs, achieving 3.5x inference acceleration while maintaining output quality at a given sparsity level. The scheme employs custom JAX and Pallas Splash attention kernels to convert theoretical sparsity into physical sparsity on hardware.
Baseten BlogBaseten collaborates with OpenAI to integrate support for open-source models into the agentic workflows of OpenAI customers, enabling organizations to optimize agentic workflows across open and closed models and route tasks to the best model. Baseten provides features such as Day zero model access, full-stack performance optimization, multi-cloud capacity and reliability, developer experience, and enterprise governance to support fast and reliable inference and seamless integration into enterprise toolchains.
Microsoft ResearchMicrosoft Research has released Quine, an AI research platform that combines a multi-modal biological world model with an interactive connection system, used to integrate models, scientific tools, literature, and researchers, and has been used to predict compounds driving tumor state changes and validate candidates; the system is designed for research, not for clinical or medical use, and its output may be incomplete or require verification by researchers, with future plans to expand access through Microsoft Discovery.
Claude.dev BlogClaude Sonnet 5.5 is the second model in the Claude 5.5 family after Opus 5.5, more intelligent, efficient, and 30% faster than Sonnet 5, but the price per token remains unchanged, as the cost is reduced due to a maximum of 30% reduction in token consumption for typical tasks.
Xiaomi MiMo · Hugging Face new modelsXiaomi MiMo team releases MiMo-V2.6-Flash-MOPD model, integrating multiple domain expert teachers to enhance reliable training and validation capabilities in long-term game development, research, and embodied intelligence, and effectively alleviates the issue of repetitive tool calls.
Xiaomi MiMo · Hugging Face new modelsXiaomiMiMo/MiMo-V2.6-Pro-MOPD integrates multiple domain expert teachers, expanding to long-term game development, scientific research, and other fields where reliable training and validation are difficult.
Anthropic ResearchAnthropic's Claude has successfully calculated the nine-loop amplitude angles for N=4 super Yang-Mills theory, breaking the previous two-loop calculation limit. The model is based on LLM technology and uses the CoT reasoning method.
Google DeepMindGemini 3.8 Live with Live Avatar released, enhancing near real-time visual presentation capabilities for Gemini conversational AI by natively coupling real-time video generation with speech, supporting 97 languages simultaneously, and available in Gemini Enterprise.
Anthropic ResearchIn the Project Swap experiment, 201 Anthropic employees conducted book trades through the Claude agent, with 61% of the agents accurately matching employee preferences. The agents performed well in the trades, but their efficiency was limited by insufficient information. The experiment showed that model performance had a greater impact on the negotiation results of the agents than instructions, with most participants satisfied with the books they received from the trades and willing to pay about one-third of their annual budget for Claude to handle the trades.
Liquid AI Blog7575Liquid AI has released the DSpark Visual Language Model Accelerator LFM2.5-VL-3B, which boosts decoding throughput by 2.66x on GPUs and 3.13x on edge devices through a lightweight drafter model.
Microsoft ResearchOffloading physical AI reasoning to edge or cloud GPUs can significantly improve robot task success rates, support larger models, and extend battery life. In mobile operation tasks, offloading reasoning enhances robot response speed and accuracy, increases battery endurance, and reduces costs. Microsoft Research has developed a Kubernetes-based toolkit to support distributed reasoning deployment on robots, edge devices, and the cloud.
Liquid AI BlogLiquid AI announced at the Snapdragon Summit 2026 that its device-side context layer, Liquid Context, has been optimized for Snapdragon® processors, specifically using Qualcomm® Hexagon™ NPU.
Xiaomi MiMo · Hugging Face new modelsXiaomi MiMo released the 9B parameter intelligent agent model MiMo-V2.6-Distill-Qwen-9B, which covers coding, general intelligent agent tasks, visual encoding, and cybersecurity fields by supervised fine-tuning of Qwen3.5-9B on data generated by MiMo.
StepFun · Hugging Face new modelsJieyue Xingchen releases flagship model Step-5-Preview-BF16, with 600B total parameters, 27B active parameters, supporting up to 1M context windows, excels in software engineering, professional knowledge work, and financial fields, capable of processing text, images, and video, deployable via OpenAI-compatible API or offline generation.
StepFun · Hugging Face new modelsStep-5-Preview, the flagship model from Jieyue Xingchen, adopts a sparse MoE architecture with 600B total parameters and 27B active parameters, supporting up to 1M tokens in context, and excels in software engineering, professional knowledge, and financial tasks, demonstrating outstanding performance in the DeepSWE v1.1 benchmark. The model can be deployed via OpenAI-compatible APIs using vLLM or SGLang, or used for offline generation with Transformers.
Google DeepMindGoogle DeepMind has launched the Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking large language models, with the former designed specifically for scale and cost-effectiveness.
Google · Hugging Face new modelsGoogle has released EmbeddingGemma 2, a multimodal embedding model with 740M parameters, supporting unified 768-dimensional vector representations for text, images, video, and audio. The model can run on consumer-grade hardware, providing low-latency semantic representations suitable for applications such as search, retrieval-augmented generation, etc. EmbeddingGemma 2 has performed excellently in multiple benchmark tests, such as text, code, visual document, and audio embedding benchmark tests, and supports vector truncation to reduce storage requirements.
Cognition BlogDevin Desktop and CLI are now integrated with the Fusion architecture, improving efficiency by up to 39% compared to other code generation frameworks. Fusion allows users to select a cutting-edge model for planning and review ('lead') and a more cost-effective model for execution ('sidekick').
Cognition BlogSWE-2 is our most advanced coding model to date, achieving a score of 50.0% in the FrontierCode 1.1 Main evaluation, approaching the performance of Fable 5.1 with a 64% cost advantage, and pioneering the application of reinforcement learning to trillion-parameter scale.
DeepSeek · Hugging Face new modelsDeepSeek-V4.1-Flash is a 552B parameter multimodal MoE model that supports context up to one million tokens, natively processing images and text and generating text autoregressively. It adopts a Causal Encoder-Decoder architecture, reducing global KV cache usage to 890 bytes/token through technologies like CSA2 and FP4 primary KV cache, a 75% reduction compared to DeepSeek-V4-Flash.
OpenHands BlogWhen evaluating open-source large language models, it's necessary to distinguish between true open-source and open-weight models due to significant differences in weight disclosure, code, training data, and usage rights. Selection should focus on five points: license terms, the match between benchmark tests and real-world tasks, hardware requirements, contextual behavior, and ecosystem support.
Zhipu Z.ai · Hugging Face new modelsZhipu introduces GLM-5.3-Flash, the first native multimodal GLM-5 series, with 320B total parameters and only 18B active parameters, benchmark tests and real task performance surpass GLM-5.2, priced only one-tenth, coding and agent testing close to Claude Opus 4.8
Zhipu Z.ai · Hugging Face new modelsCompared to GLM-5.2, Zhipu GLM-5.3-BF16 improves coding ability by 50% on Z.ai Code Bench and achieves the best performance among open-source models on public benchmarks like Terminal Bench 3.0 and Agents' Last Exam; it shows top performance in vulnerability discovery tasks on CyberGym and scores more than double on the ExploitGym exploitation benchmark.
Zhipu Z.ai · Hugging Face new modelsZhipu releases GLM-5.3-Flash, the first native multimodal GLM-5 series model with 320B total parameters and 18B active parameters, surpassing GLM-5.2 in benchmark tests and real-world workloads, priced only one-tenth of GLM-5.2, with coding and agent benchmark tests close to Claude Opus 4.8.
