OpenRouter BlogRunning the same tasks on OpenRouter, GPT-6.1 Sol passed 180/180 tests in 60 synthetic tasks, DeepSeek had the lowest cost passing 157/180 times, but the correlation with the public ranking is low (Spearman correlation coefficient 0.109). Need to check the number of task repetitions, the number of answers, and billing verification.
AWS Machine Learning BlogDevelopers use OpenAI Codex for code exploration and engineering work through Amazon Bedrock. Amazon Bedrock supports Codex inference and can be integrated with AWS Identity, IAM, and observability practices. It is recommended to start with Native AWS Access for small-scale pilots, use GPT-5.6 Sol for validation, authenticate through IAM Identity Center, and configure model permissions and observability targets.
Google launches the Gemini Decisions API, driven by the Gemini Decision1 model, allowing developers to evaluate up to 1000 questions for multimodal input, scoring preset options in a single forward pass and returning probabilities and confidence levels, with end-to-end latency of 150–250 milliseconds, and a fee of 0.10 USD per million input tokens.
Simon WillisonSame story · 7 outletsGoogle has released the multimodal decision-making model Gemini Decisions API, supporting text, image, audio, video, and PDF, similar to Jev and OpenAI's decision models but more confident; the price is 10 cents per million input tokens, output is free, more expensive than Jev (4.2 cents per million tokens) and OpenAI (same price)
MarkTechPostFeaturedSame story · 7 outlets7676
The DecoderFeaturedSame story · 7 outlets7878Google DeepMind launches Gemini Decisions API, targeting yes/no, multiple choice, and scale scoring for closed-form judgments. The underlying model, Gemini Decision 1, can handle up to 1000 independent questions at once, covering text, images, audio, video, and PDF.
Google Research5858Google Research releases a wearable device-driven CKM risk assessment framework, predicting future health trajectories through an explainable health age model (CHA). The framework is trained on wearable data from 2.2 million weeks using the SensorWM model, providing personalized 12-week counterfactual health simulations combined with language models to generate clinically validated guidance plans. The CHA model integrates modifiable indicators such as heart rate, steps, and sleep to predict disease risk over 68,000 person-years, improving the predictive performance of clinical indicators.
Apple Machine Learning ResearchState Space Models (SSMs) lag behind Transformers in contextual learning and precise retrieval tasks, as their gating mechanism leads them to first learn memory solutions, delaying or preventing the convergence of correct contextual learning solutions, but the gating mechanism is beneficial for improving long sequence generalization.
Google AI BlogFeaturedSame story · 7 outlets7878Why it matters: The original text provides forward scoring, latency, and pricing data, allowing readers to judge its suitability for which low-latency decision-making scenarios.
AWS Machine Learning BlogAWS Quick's AI-driven product development lifecycle (AI-PDLC) transforms enterprise signals into validated prototypes through six specialized skill agents, combined with human approval. The Discovery agent analyzes pain points in enterprise data sources, and the Prioritize agent uses a 0-10 score framework to evaluate use case priorities, with all decisions requiring manual confirmation.
Databricks Blog5858Databricks releases Genie Ontology, an enterprise context layer, combining managed definitions with continuous learning of organizational knowledge to provide business understanding capabilities for AI agents. Genie plugins support ChatGPT, Claude, Microsoft Copilot Cowork, and Grok Bot, enabling teams to maintain business context in Databricks and use it across multiple agents.
Databricks Blog5858Databricks Genie Agents can now perform high-quality business analysis, analyze large volumes of documents, and connect enterprise tools to automate actions by supporting unstructured data analysis and new MCP connectors. The new OpenSharing feature allows users to share managed agents with pre-installed data, instructions, and knowledge with partners, suppliers, and customers, accelerating the generation of external insights.
Hugging Face Blog3838Hugging Face's blog introduces the Token Pooling perception training method for multi-vector embedding models, which compresses storage by merging similar tokens through HierarchicalTokenPooling.
The DecoderThe DeepL Aure model conveys the speaker's tone, emotion, and pace to the target language by analyzing the original audio and translated text, supporting eight languages including English, German, and Spanish, and is suitable for DeepL Voice online meetings, face-to-face conversations, and the DeepL API.
vLLM Blog5858vLLM extends multi-modal service capabilities through the EPD (Encoder/Prefill/Decode) separation technology, supporting independently deployed visual encoders. vLLM v0.32.0 includes MiMo and XPU support, and the metadata-only request feature of DeepSeek V4.1 Flash has been open-sourced.
Ant Group inclusionAI · Hugging Face new modelsinclusionAI has released the ZoomGain-4B large language model, achieving a score of 61.08 on ZoomGainBench and performing excellently on multiple fine-grained perception and general perception benchmark tests; the model is built with supervision based on the global-to-local visual perception gap mined from global image details, supporting English and Chinese Q&A, and can be downloaded via Hugging Face.
Ant Group inclusionAI · Hugging Face new modelsinclusionAI/ZoomGain-4B-GGUF is a 175B parameter multimodal model supporting image input, using FP16, Q8_0, and Q4_K_M precision quantization. The model includes a language model and mmproj visual encoder/projector, converted to GGUF format via llama.cpp. It provides llama-mtmd-cli for command-line inference and llama-server for network chat functionality.
Ant Group inclusionAI · Hugging Face new modelsThe GGUF format weights for inclusionAI/ZoomGain-9B have been released, converted using llama.cpp, with the language model in FP16, Q8_0, Q4_K_M. The visual encoder and projector (mmproj) also adopt this quantization scheme. Both the language model and mmproj files need to be provided for image input, supporting CLI inference (llama-mtmd-cli) and Web Chat (llama-server) deployment.
Hugging Face Daily Papers6565SWAM extends the World Action Model through cross-modal alignment, trained on 47,000 hours of multimodal data, blending egocentric video and UMI demonstrations to enhance control generalization, ultimately achieving a 40.00% success rate on three real-world tasks, surpassing the baseline of 23.33%.
MarkTechPostMBZUAI's IFM releases PAN 2, an open 29B parameter diffusion Transformer (MMDiT) world model, which takes input text and optional first frame, outputs 704x1248 resolution 24 fps video (up to 20 seconds long), achieving a 34.0% complete plan success rate in 100-step multi-tasking, leading 6 test models, but with high failure rates for long plans and high hardware requirements, the code is open-sourced under Apache 2.0.
Hugging Face Daily Papers6565The DexForesight framework implicitly learns spatial foresight by injecting a dynamic perception world representation into the latent space of dexterous strategies, increasing the success rate of basic strategies from 49.4% to 60.5% on 11 DexJoCo simulation tasks and achieving a 58.3% success rate on 3 real robotic dexterous operation tasks.
Hugging Face Daily Papers6565Infrathon conducted 150 long-term AI infrastructure optimization task evaluations on six models, consuming a total of 10,000 GPU hours, and found that the model rankings changed with the evaluation duration. 135 tasks achieved further optimization in the latter half of the budget, and Claude Opus 5.5 increased the KV cache compression rate from 6.82% to 19.86%, exceeding human design baselines. Analysis shows that task violations, unreliable testing, and premature completion are the main obstacles to continuous optimization.
Ant Group inclusionAI · Hugging Face new modelsSing VL S1 is a multimodal decision-making model based on Qwen3.5, capable of answering named entity questions with text or JSON and up to four images, scoring options for each question using a single Transformers forward pass, performing excellently on 43 text datasets and 11 visual datasets, supporting Python 3.12 and PyTorch inference.
Latent Space6565OpenAI introduced new ways of working for Codex at the AI Engineers' Code Summit, including improvements to cloud functions, voice interaction, and agent collaboration. Codex Cloud 2.0 allows users to run multiple tasks in isolated instances, but there are still usability issues with cloud functions. The voice mode, as a control plane, can manage code repositories, and personal agents dots can serve as safety nets, delegation nodes, and ubiquitous presences.
Hugging Face BlogTether AI Research has launched the 0.8B parameter BotPsy model, built based on Qwen3.5-0.8B, specifically designed for function calls and agent workflows, suitable for on-device deployment. BotPsy outperforms all ≤1.2B parameter models in 6 tool use and agent benchmark tests, including Qwen3.5-2B-Think, achieving the highest scores in 5 tests, with a 2.5x speed increase and a 3x improvement in reasoning efficiency.
Hugging Face Daily PapersdNaViT is a discrete native resolution vision Transformer that matches or exceeds specific baselines in visual understanding, generation, and editing tasks through grouped product quantization and dual-branch targets. It performs excellently on 12 visual understanding benchmarks, 4 generation benchmarks, and one image editing benchmark, proving that discrete vocabulary can be a universal visual representation
NVIDIA BlogThe NVIDIA Jetson Orin NX module is used for inference on the pilot line of AIRECO's recycling plant, driving a robotic arm to pick up target objects from a conveyor belt with an accuracy rate of 95%; plans are to increase the annual processing capacity from 18,000 tons to 48,000 tons next year, reducing operating costs by 85%.
AWS Machine Learning BlogMoonshot AI's Kimi K3.1 Pro Highspeed has been released on Amazon Bedrock, the model being a post-training version of the Kimi K3.
Comfy BlogSeedream 5.5 is provided through partner nodes in ComfyUI, a multimodal model that generates content from text and edits from images. This version advances in transparent RGBA generation, precise local multi-round editing, poster and chart generation with readable text and structured layouts, more realistic portraits, and lighting control. In ComfyUI, all features are integrated into a single node, allowing for sequential editing processes and direct transparent output to synthesis nodes.
OpenRouter BlogThe GitHub MCP server added 19,436 input tokens for each request of Claude Sonnet 5.5, far exceeding the CLI (1,421) and skills (3,315).
OpenRouter BlogTesting on OpenRouter, fp8, fp4, and fp16 showed no precision differences in 240 scoring tasks, with all six endpoint scores ranging from 98.3% to 99.2%. Lower precision is generally cheaper, but there were 7 models where fp8 was more expensive than fp16, and 7 models where 4-bit was more expensive than 8-bit. The fp4 endpoint performed the worst among 40 workflows without multiplication tasks, but the difference might be due to chance.
Google launches VLA robot model Gemini Robotics 2.5, achieving leaps in dexterity, reliability, and safety, with a success rate of 96% on the Mercedes production line. The model has tested 200 tasks, with a success rate 30%-100% higher than the baseline, a speed increase of 20%-30%, and is compatible with robots like Apptronik Apollo 2, adopting ASIMOV-Jailbreak benchmark to enhance jailbreak resistance to 99.5%.
Anthropic introduces Claude Whiteboards, Claude Sheets, and Claude Data, moving AI assistants from chat Q&A to team collaboration workstations
Google, in collaboration with Artificial Analysis, has launched the MKA benchmark test to evaluate the ability of AI agents to handle multimodal knowledge work. Gemini 4 Carbon ranked second in the test, with task costs being one-sixth of the first-place cost.
AWS Machine Learning Blog4040Amazon Bedrock AgentCore Evaluations provides three evaluators—built-in, DeepEval, and AutoEval—to assess agents. The built-in evaluator is developed and tested by AWS, covering common dimensions; DeepEval and AutoEval are third-party open-source libraries, with AgentCore hosting them in production environments. The evaluation process requires first defining business success criteria, breaking them down into observable behaviors, and then mapping them to the appropriate evaluator level.
OpenAI Alignment BlogGPT-6 Astra shows a decrease in CoT monitoring ability in non-adversarial settings, an increase in CoT evasion ability in adversarial tests, and no significant improvement in CoT capability. During pre-training, CoT capability does not increase with the decrease of the base model's perplexity, and the architecture has a limited impact. Reinforcement learning enhances CoT control ability, and model-generated data fine-tuning can transfer CoT control ability. Model perplexity is positively correlated with CoT capability and control ability.
The DecoderNvidia's GR00T 2 AI model learns new tasks for robots by analyzing human video actions without remote control, trained on 400,000 hours of video data, with lab test success rates of 10-80%, and combined with GPT-6 Astra achieves 97% success rate for toy model assembly, but tactile processing and language model latency remain challenges.
