Researchers trained a 27B-parameter agent on a scalable RL environment built from paper replication; it outperforms far larger frontier models on held-out replication tasks, arguing long-horizon scientific capability lives in the weights.
Retail and institutional investors scrambled for shares in one of China's best-known humanoid robot makers, underscoring runaway demand for humanoid hardware exposure.
Per WSJ, Nvidia revised its planned financial backing for OpenAI's giant Ohio data center, cutting its guarantee from roughly $250 billion to under $120 billion to address investor concerns about risk.
Meta pushed frontier LLM judges across 14 tasks: verdicts flip 25-71% under static pushback and 62-91% under an adversarial persuader, casting doubt on AI-as-judge pipelines without robustness checks.
Trained from scratch on 161 datasets using only permissible post-training data, the 1B model outperforms HRM-Text 1B and competes with larger frontier models, setting a new state of the art for Danish.
A series of scientific agentic foundation models supporting multimodal understanding, reasoning, generation and long-horizon tasks, trained over rendered documents, interleaved image-text data and diverse scientific corpora.
A humanoid motion tracking benchmark whose evaluation aligns with human perception, targeting contact-rich long-horizon behaviors — foot skating, mistimed touch-downs — that kinematic per-frame errors miss, at larger scale than existing suites.
A benchmark that evaluates interactive video world models through long-horizon agent-player objectives, checking whether simulated environments stay consistent (360-degree turns, water ripples) instead of relying on short human clips.
A benchmark with 56 one-shot tasks from 14 incident-derived families that isolates command-generation errors from execution-transport failures in LLM coding agents, showing matched execution scores hide parser-induced failures.
Qwen released Qwen3.8-2.4T-A95B, the first open-weights MoE at 2.4T total parameters (95B active) with a 1M-token context window and strong coding/agentic performance — now available on Together AI as the open base for Qwen3.8-Max.
LMSYS added day-0 SGLang support for the compact Qwen3.8-27B dense vision-language model, claiming near-instant TTFT and >200 tokens/s decode on RTX 5090-class GPUs and easing local agent deployment.
Frontier models keep growing while small models still dominate real-world usage and local inference; Qwen leads local inference ahead of Gemma, and AI agents are becoming a major presence on the Hub.
The new /loop 5m <prompt> command re-runs a prompt on a schedule with intelligent back-off, while Hermes Desktop can now link to a persistent Hermes Cloud agent that keeps running after the laptop closes.
Suno's platform becomes a generative DAW that runs entirely in the browser, letting users compose and produce music without leaving the tab.
The post-trained model leads CyberGym and AutomationBench, with open weights staged rather than immediate.
The harness treats every component as a plugin for flexible, composable agent setups.
Free coding agents positioned directly against the paid incumbents.
Italian startups attracted $652M in VC in H1 2026 as the tech sector reached €140B; Italian AI companies raised $204M, growing enterprise value to $8.2B — nearly triple 2022 levels — and supporting about 25,000 jobs, with rounds from Lexroom, MDOTM, Smartness and others.
€393M went into regional startups during H1 2026 as Italian startup investments grew 82% year-over-year, with AI cited as the main driver.
AI code review used by 150K open-source projects and 17K customers; round led by Atomico and Smash Capital.
Granite Asia led the round for the AI-native drug discovery platform based in Shanghai and California; funds train the model via lab-in-the-loop feedback.
Andreessen Horowitz leads the round for the independent model evaluation platform; existing backers 8VC, Pear VC and Bloomberg Beta joined by HRT Ventures and Next Ladder Ventures.
Source:
