Prime Intellect Blogprime-rl v0.10.0 has been released, defaulting to the new RL algorithm IPO to improve the stability of asynchronous RL and optimize performance. It supports new models such as Qwen3.8 Flash, and increases training throughput by 1.15 times (based on the torchtitan platform) through MoE optimization and custom kernels.
Together AI BlogTogether AI has launched Post-Training services, supporting the construction of custom models through reinforcement learning and supervised fine-tuning without complex infrastructure. The service provides flexible training loop definition capabilities and optimizes the underlying system and infrastructure to improve model quality and training speed. It supports mainstream open-source models such as Muse Spark Open, GLM 5.3, Kimi K3, and Qwen 3.8-27B, running on dedicated GPU resources to ensure efficient and numerically consistent training.
MIT News · AI7575The multimodal AI model developed by MIT researchers combines 3D atomic structures with natural language descriptions, generating rich structural language descriptions from experimental research papers and first-principles calculation databases, helping AI learn the key properties of materials. The model outperforms text-only models in identifying solid-state lithium-ion battery electrolyte materials and can learn geometric structure representations, providing new research tools for materials science.
Tencent Hunyuan · Hugging Face new modelsQwen3-8B-DSpine-b16 released by Qwen, employing speculative decoding technology with parallel drafting and deep conditional injection, achieving an average acceptance length score of 10.55 on Qwen3-8B. The model supports a context length of 8192 tokens, uses bfloat16 data type and fa3 attention backend, installation requires git clone and pip install, and runs via sglang serve.
Modal BlogModal releases VM Sandboxes virtual machine sandbox, supporting complex Agent workloads, adopting a block device solution to solve the 99µs chown overhead and 6 times host system call problem added by virtio-fs, achieving sub-second cold start performance same as containers.
Tencent Hunyuan · Hugging Face new modelsHy-Memory-30B-A3B-FP8 released by Hunyuan is a compact long-term memory writing model with 3B active parameters, designed for efficient vLLM deployment, and performs better than Claude Opus 5 and DeepSeek V4.1 Flash in the cross harness macro average score of a fixed 20% benchmark matrix.
Ollama BlogReflection's first open-source weight model Beam has been fully launched this morning. Ollama is the first partner to provide this model for a limited time at no cost, and its Beam endpoint supports full zero-data retention (ZDR).
Microsoft AI NewsMicrosoft AI has completed the pre-training of a new MoE large model MAI-Base-2, with 80 billion active parameters and 2.8 trillion total parameters, surpassing all open-source models in internal evaluations. The model is 4-12 times more efficient in pre-training FLOPs than the strongest open-source base model DeepSeek-V4-Pro-Base and is used to train larger models. Microsoft plans to license the model weights through Foundry using organic data pre-training and full-stack control.
NVIDIA BlogNVIDIA releases the open-source 120B parameter Nemotron 3.5 Super VL model, achieving the fastest task completion speed and highest accuracy in multimodal agent tasks; simultaneously launching Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB edge models.
Latent SpaceLeash Bio optimizes DNA encoding library (DELs) data collection by attaching dozens of barcodes on molecules, its AI agents are on par with Boltz on public benchmarks, and has released a verifiable scientific reasoning AI tool named Minos and an open-source dataset.
AI as Normal TechnologyBeyond Automation: A new research direction in AI argues that AI should not be limited to automating human labor but should explore new directions for enhancing humans and promoting human-machine collaboration. This perspective advocates for AI research to focus on new issues such as transparent agents, process tracking, and human-machine interaction design to push AI beyond pure technological automation and achieve broader applications.
Mistral AI · Hugging Face new modelsMistral AI releases Timestral-123M time series base model, with 123M parameters, using a patch-based Transformer architecture, context length of 32,768 time steps, single prediction covers 1,024 time steps, outputs 21 quantiles. The model achieves optimal accuracy in zero-shot prediction on GIFT-Eval, TIME, and fev-bench evaluations, and is licensed under the Modified MIT License.
Mistral AI · Hugging Face new modelsStepFun · Hugging Face new modelsLeapX Stars releases flagship model Step-5-Preview-NVFP4, with 600B total parameters and 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.
Google DeepMindGemini 4 Argon ranked first on the Vals Index, surpassing other cutting-edge models in coding, finance, law, cybersecurity, and science with its 175B parameter size; its reasoning depth and accuracy make it a powerful tool for solving complex engineering problems, optimizing business processes, and accelerating scientific breakthroughs.
LangChain BlogScaleDown optimizes context management for long-running agents through compression and extraction techniques, its model is 41 times cheaper and 2.7 times faster than general models like GPT-5.5 in benchmark tests. ScaleDown provides four models: Summarize, Compress, Extract, and Classify, respectively for summarization, compression, extraction, and classification, supporting up to 1M tokens of input, and supporting image and PDF input.
MIT Technology Review · AI7070Tech companies like SimpleClosure and Sunset are buying code and workspace data from failed startups to sell to AI labs, as AI training labs face limitations with public and synthetic datasets.
Microsoft ResearchMicrosoft Research has open-sourced the multimodal model StudentSim for simulating student learning behavior. It is trained on publicly anonymized learning records to create a personalized student simulator that outperforms benchmarks in areas such as chess, English writing, and mathematics. It provides early feedback signals for AI tutoring systems but is not intended to replace teachers or classroom teaching.
Import AI4242Humanity’s Sixth Sense review shows GPT 6 Astra scores 53.6 in intuitive visual reasoning tasks with a pass@1 rate, far below the human baseline of 93.1, with the open-source Chinese model GLM-5.3-flash scoring 26.8 and Kimi K3 scoring 25.5.
The Decoder7575Anthropic has introduced two new beta features for Claude: Dashboards can generate automatically updated real-time dashboards by connecting to data sources like BigQuery and Databricks via text prompts, with results outputtable to tools like Amplitude and Grafana; Motion can create editable code animation explanation videos from text, charts, and images and export them as MP4.
MIT Technology Review · AI6565Tesla's Optimus humanoid robot is expected to be sold to the public by the end of 2027, with each priced at about $20,000, but experts like Yann LeCun question whether its intelligence level will reach practical levels.
Google ResearchGoogle Earth AI's Population Dynamics Foundation Model (PDFM) compresses search trends, population flows, building environment density, and environmental factors into location embeddings, enhancing epidemiological research performance in various disease fields and geographical environments without specific task fine-tuning.
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.
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.
MIT Technology Review · AI6565Enterprise AI Agent Projects Average Only 34% Deployment, Knowledge Deficiency as Main Bottleneck. Knowledge-capable production leaders (61% projects passed pilot) show significant advantages in semantic understanding, but data fragmentation (55% consider it the primary challenge) and traditional system limitations remain key obstacles. Enterprises will prioritize strengthening the structural connection between data and agents through knowledge graphs, RAG technology, and AI evaluation agents.
Import AI7575Toby Ord's analysis suggests that AI population expansion accelerates task completion due to parallel processing, but adding 10 times more agents only improves performance 3-5 times, creating a 'stampede effect' similar to human team collaboration; 61% of Americans distrust corporate AI self-regulation.
Google ResearchGoogle Research announced a new generation of a federated learning system based on trusted execution environments (TEE), achieving fully verifiable and auditable data anonymization by executing encrypted training data within the TEE and publishing access policies to the public transparent log Rekor, Gboard has adopted this system and significantly improved computational speed.
NVIDIA BlogThe NVIDIA DGX Spark with 64GB unified memory, launched by Acer, ASUS, Dell, Gigabyte, HP, and MSI, supports developers, researchers, and AI enthusiasts to run models with up to 100 billion parameters locally without cloud dependency. Two devices can be clustered via NVIDIA Sync Cluster Assistant, merging memory to 128GB and boosting performance 1.7 times, priced from $4999.
Baseten LabsQwen2.5-1.5B achieved 66.7% accuracy (63.8%) on the MATH benchmark through a 1000x single gradient step length extension to 500 steps for reinforcement learning models, with the main improvement coming from fixing issues with the base model's cyclic output, but performance degrades on out-of-distribution tasks such as function calls and Knights & Knaves.
Anthropic ResearchFeatured7171Anthropic has released a Robot Exposure Index study, using Claude to evaluate about 19,000 tasks in O*NET, finding that existing robots can complete 74% of physical tasks in the U.S. and account for 34% of all work time, but most are limited to controlled environments.
Why it matters: Anthropic used Claude to rate O*NET tasks item by item, quantifying how much work robots can currently cover, and provided constraints on cost and regulation.
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.
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.
Anthropic ResearchFeatured8080Anthropic has established a life science research team and built its own laboratory, allowing the Claude intelligent entity to autonomously search in DNA sequence databases and discover a new enzyme system ART (array-associated reverse transcriptases) related to DNA repetitive sequences.
Why it matters: Anthropic's self-built lab enables Claude to autonomously screen DNA databases and discover a new enzyme system similar to CRISPR, demonstrating the specific process of AI-driven hypothesis generation.
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.
Hamel Husain7575Collect at least 50 passing and 50 failing human-labeled examples through error analysis to identify discrepancies between LLM judge and human labels for repair clues. Common issues include missing context or ambiguous instructions. Manually inspect a few inconsistencies first, then attempt automated prompt adjustment, such as using the GEPA algorithm to adjust judge prompts and measure their consistency with human labels. The most common error is letting the LLM judge capture too many types of errors at once. It is recommended to create separate judges for each failure type.
Google DeepMindGoogle DeepMind releases AlphaGenome Atlas platform, containing predictions of 9 billion single nucleotide variations in the human genome, providing a free website access for academic research
Microsoft ResearchGigaPath-Flash and GigaTIME-Flash significantly reduce computational requirements through streamlined ViT-S encoders, making large-scale pathological research more feasible while maintaining performance.
Anthropic Research5858Anthropic allows external research institutions to analyze 250,000 Claude conversation data through Anthropic Insights, which is the first time external researchers can independently study the data used by an AI company itself.
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
