10 results on this page
AntLing has released six official Ling-3.0 base checkpoints spanning two model sizes and three training stages. The MIT-licensed, ungated repositories are intended for continued pretraining, fine-tuning, and research rather than direct chat use.
The post argues that Qwen3.8 intermediate tokens should not be interpreted as human-like reasoning. Citing research, it highlights that trace validity often does not correlate with answer correctness, and that models trained on corrupted or irrelevant traces can perform as well as or better than models trained on valid traces.
Tactile robotics is expanding from sensor hardware into datasets, foundation models, and evaluation benchmarks. Research shows touch improves physical reasoning and robot success, but inconsistent sensor formats create major cross-device transfer challenges.
Nokia is reportedly planning further layoffs, facility closures, and a major reduction of its mainland China operations, potentially retaining only limited after-sales services. The retreat is part of a global restructuring toward AI and cloud businesses, but it may reduce Nokia’s access to China’s large-scale 5G deployment data needed for AI-RAN and future 6G research.
A hyper-optimized Qwen3.8-27B inference stack using DFlash2 reaches about 138 tokens per second on a power-limited RTX 3090, while 64-request throughput reaches 942 TPS. Prefix caching also reduces long-chat follow-up latency from roughly 23 seconds to under 1.4 seconds in reported tests.
Unitree has launched a biomimetic seven-axis dexterous robotic arm, priced from 9,900 yuan. The product could lower the hardware cost of embodied AI development, robotic manipulation research, and automation prototyping.
VB Pulse data shows enterprises are adopting multiple AI orchestration platforms instead of relying on a single vendor. The survey also highlights persistent concerns about security, permissions, token usage, cost visibility, and the ability to stop runaway agent spending in real time.
Tencent's Hunyuan Hy4 has reportedly appeared in the model selection list of the Yuanbao app under an expert-level label and with tool-use capabilities. It is positioned above Hy3 and alongside DeepSeek, following Tencent's recent statement that a larger-parameter Hy4 would launch soon with improved performance and multimodal abilities.
An independent builder pretrained a 1.02-billion-parameter Kimi-K3 replica on 5 billion decontaminated tokens for roughly $250. The model achieved 33.4% on HellaSwag, exceeding the reported 28% score of GPT-2 124M despite receiving no instruction tuning.
Meitu Imaging & Vision Research Institute (MT Lab) proposes a new scene text editing approach for languages ranging from Chinese to low-resource languages. The research is presented as an ICML 2026 contribution and focuses on making text changes appear seamless within images.