IFM has open-sourced six models in the K2 Horizon series, covering scales of 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B, along with releasing data recipes, training code, intermediate checkpoints, fine-grained logs, evaluation results, and the xLLM infrastructure. The 36B-A4B model combines MoE and MoVA sparse attention, activating approximately 4B parameters per token, achieving performance close to a dense 32B model. After reward cheating audits, the Terminal-Bench 2.1 score of the 375B model was revised from 70.2% to 66.9%.

The Mostik team introduced a cross-model bridging technique that enables direct transmission of hidden states between models, rather than compressing to text and reconstructing semantics at the receiving end. In experiments, the cloud-based 753B GLM-5.2 only performed prefilling, then passed internal representations to the mobile 4B Qwen-3.5 via a small bridge network. Both models' weights remained frozen, with only the bridge component trained. In the ARC-AGI 3 demonstration, this approach increased accuracy by 25%, closed about half of the performance gap, achieved up to 2x improvement on difficult subsets, and reduced large model inference costs to approximately one-twentieth.

Google partnered with Cathay Pacific to conduct AI-powered contrail avoidance trials on ultra-long-haul flights across the Asia-Pacific region. The system integrates AI predictions, satellite imagery, and weather data to identify airspace prone to forming warming contrails in advance, then provides minor altitude adjustment recommendations to dispatchers and pilots via onboard networks and electronic flight bags. Initial trials involving over 80 flights estimated a ~40% reduction in contrail-induced warming impact, with the Hong Kong to Singapore route contributing over half of the emission reductions. Google notes that contrails account for about one-third of aviation's climate impact and plans to expand into a second-phase trial.

A new study evaluated the moral reasoning capabilities of nine leading models across three deployment scenarios, requiring systems to satisfy four structural conditions: conclusion stability, monotonicity, decision clarity, and Pareto feasibility. Results showed no model could form consistent strategies across all scenarios. Simply altering problem phrasing—without changing underlying conditions—led to shifts in decision ratios of up to 99 percentage points within a single escalation tier. The paper concludes that current agents cannot reliably map moral features into consistent judgments, and value alignment evaluations should first verify basic policy coherence.

Microsoft has released the MAI-Image-2.6-Flash image generation model, claiming a 2.8x speed increase, over 50% cost reduction, and support for multi-image references and dynamic aspect ratio outputs. Positioned for high-concurrency, low-latency industrial applications, multi-image input allows integration of multiple visual references, while dynamic ratios reduce extra cropping and re-layout efforts across different content specifications. This update focuses not merely on expanding generation capability, but also on reducing latency and unit cost simultaneously, offering a higher-throughput model option for bulk asset creation and online applications.

A user report based on LobsterAI's million-scale real-world data shows that average task size for office agents grew 3.1x over five months, with an August month-on-month increase of 53%, indicating a shift from simple Q&A to long-cycle complex tasks. Nearly 60% of token consumption occurs outside regular working hours, and 12.2% of model calls are not initiated in real time by users. Scheduled tasks are driven by a single instruction triggering an average of 64 model calls—about six times that of manual tasks. Usage is highly concentrated: the top 20% of users consume 87.4% of compute, with general programming and debugging accounting for 38.5% of all tasks.

Alibaba's Qwen Office has launched a multi-user workspace, allowing users to describe requirements in natural language to generate and publish web applications supporting real-time collaboration for up to 100 people. The feature includes built-in role permissions, cloud databases, admin panels, and online publishing, suitable for business scenarios requiring multi-user data entry, queries, and workflow coordination, with ongoing modifications possible via natural language. Qwen Office reportedly gained over 30 million users within one month, more than half being enterprise users. This update extends agents from individual content generation to organizational collaboration workflows with access control and persistent data.

A privacy study on Split-LLM reveals that in a two-node training setup, mixing real data with decoy rows fails to prevent an untrusted cloud server from identifying original samples. Since the loss function ignores decoy data, their returned gradients are exactly zero, enabling the cloud to directly distinguish real from decoy rows using zero-value patterns, thus bypassing privacy checks already passed at the forward pass. Experiments show gradient clipping and noise addition can mitigate this specific signal, but overall security cannot be guaranteed, as attacks involving cumulative observations across multiple training steps remain untested. Privacy audits must therefore cover bidirectional data flows.

Zhixiang released the embodied world model HiDream-O1-Embodied, topping the RoboColiseum robustness leaderboard with a score of 0.692. The benchmark evaluates model generalization under non-ideal conditions by varying background, lighting, materials, and camera positions. The model employs multi-view collaborative perception and actively introduces lighting variations and image corruptions during training to reduce systemic errors caused by single-path vision failure. Its data strategy builds on high-precision motion capture data, then uses generative models to create physically plausible variant videos, achieving hundredfold-scale data augmentation.