The day's essential AI stories, from the labs, startups and researchers shaping it.
Google ResearchFeatured8888Google has opened the /teamwork multi-agent workflow in Antigravity to external researchers and developers, which can be used in conjunction with Gemini 4 Argon, Carbon, and Hydrogen; /boost and Deep Review are also available.
Why it matters: Google has disclosed /teamwork's specific results in multi-agent research and engineering tasks and made it available for external researchers to use.
vLLM BlogFeatured6060vLLM provides support for Inkling 2 Small on its first day of release. The model is a multimodal MoE with text and visual inputs, featuring a total of 594B parameters, 17B activation parameters, a maximum context length of 1M tokens, and comes with a built-in DSpark beam decoder.
Why it matters: The original text provides details of vLLM's Day-0 support for Inkling 2 Small and multiple kernel optimizations, allowing readers to understand the deployment path and throughput performance of the model.
Google AI BlogFeatured8282Google Search has launched Gemini 4 Hydrogen, the largest model upgrade to date, now available to all global users. The company claims the new model can understand multi-part questions, spelling errors, and highly specific requests, and leverages stronger instruction following and tool calling capabilities to respond to each question individually. Hydrogen's reasoning ability also enhances query distribution, allowing it to retrieve multiple related subtopics simultaneously and provide more images, structured responses, comparison tables, and custom tools.
Why it matters: Google's official statement on Gemini 4 Hydrogen explains changes in intent understanding, query distribution, and visual responses in search, which can be used to judge the direction of adjustments to the search experience.
Google DeepMindFeatured8787Google DeepMind releases Gemini 4 Hydrogen, targeting daily coding and intelligent agent workflows, with an Artificial Analysis intelligence index of 51 at the default high thinking level, higher than Argon's 50, and reaching 54 at the xhigh level
Why it matters: Google provides the division of labor and positioning of Hydrogen, Carbon, and Argon, as well as the actual differences in cost and speed between multi-model arrangements.
vLLM BlogFeatured6060The vLLM-Omni team has integrated MiniMax H3-Image into NVIDIA Vera Rubin. With a four-card SP4 configuration, the native 4096×4096 image-to-image short prompt benchmark median latency is 0.87 seconds, and the throughput of four copies SP1×DP4 reaches 6115 images per hour.
Claude · YouTubeFeatured6868Claude Mythos 5.2 predicts the complete AAV capsid structure, which is composed of 60 identical protein chains. It writes code for an open-source folding model to handle the overlap grouping of adjacent five chains, then uses the symmetry of the capsid to assemble the 60 chains. On AAV2, its prediction deviates from the experimental structure by about the width of an atom.
Why it matters: The original text provides specific methods and accuracy of the model in real scientific research tasks, allowing readers to judge its applicable boundary in structural prediction work.
Claude · YouTubeFeaturedSame story · 3 outlets7575Anthropic releases Claude Fable 5.2, designed for the most challenging technical challenges, with improvements in coding, reasoning, and vision, suitable for the most ambitious open-ended technical work. The official recommendation is to use Opus 5.5 for most coding and knowledge tasks, and Fable 5.2 is now available in the Pro, Max, Team, and Enterprise packages.
How the accounts differ:TechCrunch states the input price of Fable is reduced to $8 per million tokens; others do not mention the price.
Why it matters: The official statement explains the positioning and applicable scenarios of Fable 5.2, and provides a choice suggestion between Opus 5.5, making it easier to decide which model to use.
vLLM BlogFeatured7272NVIDIA Research and Inferact's REINFORCE Zero proposes to first align rollout with training kernel, then use unclipped token-level importance sampling to correct weight lag, achieving a training step throughput 2.40 times that of synchronous training under 16 GPUs and 12K updates.
Why it matters: The paper breaks down the gap between asynchronous RL's rollout and training into two items: weight lag and kernel inconsistency, and proposes an alignment-free, no-trimming importance sampling method after alignment.
vLLM BlogFeatured7575vLLM announces Day-0 support for the open-sourced Kimi K3.1 Pro on NVIDIA, AMD, Google TPU, and Alibaba's Zhenwu PPU, a 2.8T parameter (104B activation) multimodal MoE model with 1M context tokens.
Why it matters: The vLLM team details the engineering trade-offs for Day-0 adaptation of Kimi K3.1 Pro, including specific numbers for sparse attention, KV cache packing, and DCP.
Together AI BlogFeatured2 reports7575Together AI has released Together Compiler (TKCC), which uses high-throughput intelligent agents to continuously generate optimization solutions for inference engines, achieving 1.3-3x TPS improvements on cutting-edge open-source models like Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash within 24 hours.
How the accounts differ:Together AI Blog mentions models Kimi K3 and GLM 5.3; MarkTechPost mentions SemiAnalysis InferenceX AgentX.
Why it matters: The original text provides a comparison of throughput improvements and engine startup time on multiple cutting-edge models, allowing for an assessment of changes in the inference optimization process.
Anthropic ResearchFeatured6262Anthropic announced the verifiable reasoning prototype FlashProve, using zkSNARK to prove that the response indeed comes from the committed model weights. For the 400B parameter Llama 4 Maverick (4-bit quantization), a single H100 generates proofs at approximately 110 tokens/second, with proofs for each 1,024 token request being about 19 MB, and verification on a 24-core CPU taking about 17 seconds.
Why it matters: Anthropic has released the measured data and Lean formal proof of the verifiable reasoning prototype FlashProve, allowing for a cross-comparison with existing zkML solutions.
Hugging Face BlogFeaturedSame story · 2 outlets8585Mistral releases multimodal MoE model Mistral Large 4 (Le Chonk), with approximately 1T total parameters, 49B activation parameters, a context window of 1M tokens, offering two official versions FP8 and NVFP4, as well as GGUF made by Unsloth and llama.cpp teams.
Why it matters: The original text provides specific scores for the 1T total parameter, 49B activation multimodal MoE model on cybersecurity and coding benchmarks, along with measured throughput for local and endpoint deployment.
Google DeepMindFeatured8787Google DeepMind has opened Gemini 4 Carbon to developers, enterprises, scientists, and consumers, claiming it has reached cutting-edge levels in network security capabilities and long-range intelligent agent workflows.
Why it matters: The official provides the positioning division of roles between Carbon and Argon, multiple benchmark rankings, and cost comparisons per task to help determine when to use a heavier model.
vLLM BlogFeatured6060The vLLM team used P/D separation and Wide-EP to serve DeepSeek-V4.1-Flash on a single NVIDIA GB300 NVL72 rack, achieving 954K tokens/s per GPU and 196 tokens/s for single-user P90 on SemiAnalysis AgentX, which is 7.9 times the early single-node results.
Why it matters: The vLLM team has released per-item optimization and measured data for serving DeepSeek-V4.1-Flash on GB300 NVL72, which can be compared with DeepSeek's own engine DeepInfer.
vLLM BlogFeatured6262vLLM team releases the second part of GLM-5.3 optimization, deploying on 16 H200s with 1 pre-filled node and 1 decoding node, using PCP to split pre-fill by token and DCP to split KV cache by token, allowing 1M context to fit into the pre-fill node
Why it matters: The vLLM team has dissected how PCP, DCP, and Hybrid HiSparse can be combined on 16 H200s, allowing readers to understand the memory and pre-fetch bottlenecks of long-context deployment.
Hugging Face BlogFeatured7878NVIDIA releases pre-trained relational foundation model Kumo Relational, enabling direct prediction on enterprise databases without task-specific training, feature engineering, or table flattening. The model outperforms supervised relational deep learning and expert feature engineering gradient boosting trees on 41 tasks across 15 databases, achieving an MRR 8% higher than AutoGluon ensemble on the SAP SALT benchmark.
Why it matters: The original provides a comparison result of pre-trained relation models on 41 tasks and an open-source entry, allowing readers to judge whether it can replace the existing XGBoost pipeline.
NVIDIA BlogFeatured6262NVIDIA expands its family of open models to cover language, robotics, biology, and quantum computing, including the new Nemotron 3.5 Super VL and Nemotron 3.5 Bolt.
Why it matters: NVIDIA has rolled out open models across four lines: language, robotics, biology, and quantum, allowing readers to judge the coverage of its open-source ecosystem based on this.
Google AI BlogFeatured9090Google has released Gemini 4 Argon, opening it up to developers, businesses, and researchers, with API pricing at $2 per million input tokens and $10 per million output tokens, offering a 95% discount on cached input tokens.
Why it matters: Official pricing, baseline rankings, and multiple internal implementation cases are provided, allowing judgment on its position in coding and research workflows.
