OpenAI Launches GPT-Realtime-2.1-mini with Reasoning and Tool Calls
OpenAI released the GPT-Realtime-2.1-mini model in its API, bringing reasoning and tool-use capabilities at the same cost as GPT-Realtime-mini, suitable for real-time interaction scenarios where agents need to think before responding and call external functions mid-conversation.
Google DeepMind Partners with Apptronik — Apollo 2 Humanoid Data Trains Gemini Robotics
Apptronik expands its Robot Park flagship facility to nearly 90,000 square feet and launches the Apollo 2 humanoid in bipedal and wheeled configurations. Through its partnership with Google DeepMind, real-world operational data collected across Robot Park and customer sites such as Mercedes-Benz and GXO feeds directly into training and optimizing the Gemini Robotics foundation model, forming a "work-data-improvement" iterative loop.
I've long suspected that models have some kind of "working memory" to store intermediate variables during a forward pass and this paper has the best evidence yet.
— Neel Nanda, on Anthropic's workspace research
NVIDIA Nemotron Cited by 145 Papers at ICML 2026
At ICML 2026 in Seoul, 145 accepted papers cited NVIDIA's Nemotron model and datasets, 74 NVIDIA-authored papers were accepted, and approximately 2,000 papers referenced NVIDIA GPUs — cementing open models as the foundation for modern AI research.
Mistral Leanstral 1.5: 119B MoE Scores 100% on miniF2F, New SOTA in Mathematical Proof
Mistral AI releases Leanstral 1.5 under Apache 2.0 — a 119B total parameter Mixture-of-Experts model with only 6B active parameters. It achieves perfect scores on miniF2F, sets new SOTA on FATE-H at 87% and FATE-X at 34%, and solves 587 of 672 PutnamBench problems at roughly $4 per problem, punching far above its weight class as a Lean 4 proof agent.
Sakana AI Releases Sakana Translate: Deep Japanese-English-Chinese Translation with Cultural Nuance
Built on the Japan-adapted Namazu model, Sakana Translate offers three modes — translation, proofreading, and inquiry — supporting bidirectional Japanese, English, and Chinese with approximately 5,000-character real-time output. It preserves Japanese honorifics, cultural concepts, and internet slang that standard tools miss, approaching frontier model quality on WMT 2024 XCOMET-XL evaluation. Free to use, with file translation and API features planned.
vLLM Supports Tencent Hy3 from Day Zero with Native Tool Calls and MTP Speculative Decoding
The vLLM project announced native support for the new Hy3 Mixture-of-Experts model on launch day, including tool-call and reasoning parsers along with MTP speculative decoding, verified on both NVIDIA and AMD hardware targeting efficient agentic workflow inference.
SGLang Provides Day-Zero Hy3 Support with Verified Deployment Commands
SGLang released complete launch commands and a cookbook for Tencent's Hy3 model on release day, supporting the BF16 MoE architecture with mixed thinking, native tool calling, 256K context length, and MTP speculative decoding capabilities.
Tencent's Yao Shunyu Hails Hy3: A Leap in Reasoning, Anti-Hallucination, and Reliability
Tencent AI Lab researcher Yao Shunyu described Hy3 as a massive leap forward from Hy2 in under half a year — not just in reasoning and agentic capabilities, but especially in anti-hallucination, reliability, and product experience. More is on the way.
Hy3 Beats GLM 5.1 in Blind Tests — Makes Tencent a Leader in Open Source Models
teortaxesTex notes that Hy3, a 295B MoE with 21B active parameters, defeats GLM 5.1 in blind evaluations and may push the industry toward its base GQA architecture. If the scores hold, Tencent has become one of the premier open-source AI model providers.
ThinkingCap-Qwen3.6-27B Cuts Thinking Tokens by 50%
bottlecapai releases ThinkingCap, fine-tuned from Qwen3.6-27B using state-of-the-art algorithms on curated problem sets. It reduces average thinking tokens by 50%, with reductions exceeding 90% in best cases, dramatically improving reasoning efficiency.
SGLang Launches DSpark: Speculative Decoding That Stays Fast Under High Concurrency
DSpark optimizes variable-length verification to keep speculative decoding from slowing down in high-concurrency settings, now supporting Qwen3 and DeepSeek-V4 in SGLang for sustained inference throughput.
Vidu S1 Brings Custom Characters and Real-Time Video Calls to AI Interaction
Vidu AI launches its S1 model, allowing users to create custom characters and conduct real-time video calls — a new mode of AI interaction blending personalized avatars with conversational video. Users can bring their imagination to life with invite code VIDUS1 available for free trial.
Replit Agent Achieves Self-Improvement Loop, Says CEO Amjad Masad
Replit CEO reveals that the company's coding agent is improving rapidly through a closed self-improvement loop — the agent evaluates and refines its own outputs autonomously, marking a significant evolution in AI coding capabilities. Technical details forthcoming.
CMU's PACE: Predict Agentic Task Performance Using Simple Single-Turn Evaluations
PACE converts complex multi-step agent tasks into single-turn evaluations, enabling researchers to estimate how well an LLM will perform on agentic tasks without running slow and expensive full agent benchmarks. The method dramatically reduces evaluation time and cost while maintaining prediction accuracy.
Neuronpedia Jacobian Lens Opens Interactive Window into LLM Internal Activations
Neuronpedia releases Jacobian Lens for Qwen3.6-27B and other models, enabling interactive exploration of internal model activations in Jacobian space. The tool supports text reports, directional modulation, and multi-hop reasoning analysis modes, advancing mechanistic interpretability research with an accessible visual interface.
Developer Uses Gemini 3.5 Flash Agent to Discover 13.4-Billion-Year-Old Galaxy Candidate
Using the Antigravity Framework with a Gemini 3.5 Flash agent, a developer identified a candidate galaxy at redshift z=12.69 from JADES survey data of JWST's GOODS-South field using the Lyman-break dropout technique, demonstrating AI's growing role in astronomical discovery.
Grok Imagine Now Generates 15-Second AI Videos with Stunning Quality
xAI's Grok Imagine function now supports 15-second video generation with coherent dialogue and impressive visual fidelity. The upgrade marks a significant step forward in consumer-facing AI video creation capabilities.
François Chollet: Benchmark Scores Need Cost Efficiency, Not Just Scalars
Chollet argues that reporting benchmark results as a scalar number like "75% on XYZ" is meaningless. Researchers should always report efficiency scores, such as "75% at $10 per task," to reflect the real tradeoff between capability and cost that practitioners face.
Huawei Tau Scaling Paper Gains Renewed Attention
teortaxesTex highlights the underappreciated significance of Huawei's Tau scaling paper — not just 3D stacking but a high-design-density approach analogous to trading more pretraining for inference efficiency, boosting per-watt performance.
LongCat 1.6T MoE May Have Trained on Huawei 910B NPUs
Analysis suggests the LongCat 1.6 trillion parameter MoE model was trained on older Ascend A2 (910B) hardware rather than next-gen 910C CloudMatrix 384, highlighting breakthroughs in China's AI hardware constraints.
Runway Opens First French Office in Paris
Runway establishes a new AI research hub in Paris anchored by an initial team of ten employees, leveraging France's concentration of world-class research institutions and one of the world's strongest AI ecosystems.
OpenAI Codex Builds Tools for Biology Research
In OpenAI's Builders Unscripted series, Derya and Romain discuss using Codex for cell analysis, simulating immune cells, and the future of AI-powered scientific experiments involving simulated biology.
FactoryAI and Hugging Face Team Up for Open Agent Traces
The partnership aims to make open agent traces the fuel for training the next generation of open coding models, democratizing access to agent behavior data for community-driven model development.
NVIDIA Nemotron Family Surpasses 100 Million Downloads
The Nemotron model family crosses a major milestone with over 100 million downloads on Hugging Face, as NVIDIA AI thanks the community building with open foundation models.
LlamaIndex + LanceDB Tackle Messy Enterprise PDFs with LiteParse
Most agentic retrieval demos assume clean documents, but enterprise reality is messy PDFs. LlamaIndex partnered with LanceDB to explore LiteParse for handling tables, figures, and complex layouts in agent retrieval pipelines.
Adobe Firefly Reimagines Image and Video Generation Experience
Adobe Firefly updates its creative interface with personalized character generation in both image and video modes, with a private beta waitlist now open for users seeking main character energy.
Yann LeCun: LLMs Are Not the Path to Human-Level Intelligence
In a BBC interview, LeCun reiterated his long-held view that large language models cannot achieve human-level intelligence and that a new paradigm is needed beyond autoregressive token prediction.
STATION F Expands F/ai Startup Program with 6 New Partners
Paris-based STATION F adds Eleve Labs, Nebius and 4 other partners to its AI startup program, plus a San Francisco Week module for international teams. 80% of first cohort were repeat founders, cumulatively raising $34 million.
NVIDIA Shapes AI for Global Impact at AI for Good Summit
At the Geneva summit, NVIDIA works with cross-sector leaders to close the infrastructure gap, expand access to compute, and advance digital trust in an era of autonomous AI systems.
ICML Best Paper Winner Shares Journey: Rejected by NeurIPS, Awarded at ICML
Jared Manning's paper won one of only 7 awards out of 24,000 ICML submissions. He revealed the work had been previously rejected by NeurIPS — cleaned it up slightly and it received great reviews at ICML, encouraging researchers never to give up.
Claude Code Engineer Reveals "Unhobbling Claude" Strategy for Fable 5
At AI Engineer World's Fair, Thariq Shihipar shared how to maximize Fable 5's capabilities — the core principle is trusting the model's autonomous ability rather than constantly checking and over-fragmenting tasks.
Paper Challenges RLHF: Monotonic Inference Strategy Is the Real Optimization Target
A new paper argues that monotonic reasoning strategy — not training policy optimization — is the true objective of LLM reinforcement learning, challenging the foundations of existing RLHF frameworks.
From a Few Hundred Thousand to 100M AI Builders by 2028
Hugging Face CEO Clem Delangue's vision: AI will create tons of jobs if the technology is made distributed and open instead of concentrated in a few hands, aiming for 100 million AI builders worldwide.
Beyond Faster and Cheaper: What AI Actually Changes for Creators and Studios
Wonder Studios explores how AI is reshaping creative production beyond cost and speed — fundamentally altering workflows, creative processes, and the economics of entertainment.
Anthropic's J-Space Discovery Was Implied Since Steering Vectors, But Practically Powerful
Argues the finding was scientifically trivial but Anthropic turned it into a significant capability program tool with real practical impact.
Can a Breast Cancer H&E Slide Predict Recurrence Risk?
New research published in Nature asks whether routine pathology slides can help answer the difficult question of cancer recurrence likelihood using AI analysis.
Building the Document Context Layer for AI Agents
AI agents are the new knowledge workers, but most knowledge remains locked in unstructured documents — the race is on to build the context layer.
OpenClaw Lands on Hugging Face Local Apps
Users can now pick any GGUF or MLX model on HF and run it locally with OpenClaw's onboard setup in three simple steps.
Three "DeepSeek Moments" Expected Within Two Weeks
teortaxesTex forecasts DeepSeek V4, SuperPod 950 compute platform, and LandSpace ZQ-3 landing arriving together, potentially triggering industry-wide reaction.
Overwriting Claude's J-Workspace Could Degrade Surface Reasoning
Speculation that injecting orthogonal noise into Claude's J-workspace degrades reasoning quality even without decoding — a potential attack vector on language model inference integrity.
Opus 25 Confirmed — Community Buzz Around Mysterious Model Release
A brief mention of "Opus 25 confirmed" is sparking speculation in AI circles about a potential new model release or version upgrade. Details remain scarce.
tinyrouter: Reverse-Engineered Routing Architecture from Skana AI's Fugu
An open-source replication of the routing mechanism behind Skana AI's Fugu model, making advanced model routing techniques accessible to the community.
Replit Head of AI: People Will Use Far Fewer Apps as Agents Replace Them
The prediction foresees a future where AI agents absorb the functionality of many standalone applications, consolidating user workflows into agent-driven interfaces.
4.5-Hour CUDA + ThunderKittens GPU Programming Lesson Released
A comprehensive free educational resource covering CUDA programming and the ThunderKittens framework, created by a Stanford PhD student and TK co-author.
Local Agentic AI on Modest Hardware: Gemma 4 Meets MATLAB via LM Studio
Demonstration that 128 GB of memory is nice but not required — connecting Gemma 4 in LM Studio with MATLAB enables local agentic AI workflows on everyday hardware.
ICML 2026 Seoul + AMI × SBVA Mixer Opens Thursday Evening
Researchers gather in Seoul for ICML 2026, with the AI community mixer set for Thursday evening as the premier machine learning conference kicks off its annual gathering.