Ling-3.0-flash: Hybrid Linear Attention Balances Efficiency and Performance
Ling-3.0-flash, released by inclusionAI, is a sparse MoE model that activates only 5.1B of its 124B total parameters per token. Its defining feature is a natively hybrid linear attention architecture that stacks "Kimi Delta Attention (KDA)" and "MLA" alternately in a 5-to-1 ratio from the earliest stages of training, aiming for a leap forward in both long-context processing efficiency and computational cost. Combined with 1/64 sparsity in the MoE, the model reportedly matches or exceeds the performance of a conventional trillion-parameter-class flagship (Ring-2.6-1T) with far less compute. It also emphasizes practical utility for agentic use cases, trained through more than 10,000 interactive learning environments—suggesting a design built with production deployment in mind.
Swapping attention for linear attention, however, is not yet a silver bullet. Another study released the same day reported an experiment converting 21 of the 28 layers in Qwen3-0.6B-Base into KDA layers. Although aligning hidden states and distilling with KL divergence brought language modeling perplexity close to the teacher model, accuracy on multiple-choice questions stalled at 25–29% versus the teacher's 50.6%—near random. The results show that "invisible damage" can occur even when the apparent loss is small, silently destroying capabilities the model had presumably acquired—and they can also be read as validating Ling-3.0-flash's choice to train with a hybrid architecture from the start.
Tiny Models on the Edge: 1.5B Speech Generation on an iPhone
The push toward smaller, genuinely usable models is rapidly becoming a reality on edge devices as well. A community demonstration reported running VibeVoice 1.5B, a roughly 1.5-billion-parameter speech generation model, on an iPhone—with about 2.2GB of memory usage and generation at up to 1.28x real-time speed. The optimization was carried out through the audio.cpp inference stack, and memory usage reportedly remained stable even during long generations. A world where speech AI runs entirely on-device, without the cloud, is steadily being built out by the hands of enthusiasts.
There are also examples probing the limits of small scale. GPT-X2.5-135M, at just 135 million parameters, reportedly ranked third on Hugging Face's Open SLM Leaderboard, overtaking Meta's MobileLLM-R1-140M. It's a vivid demonstration that even with an order of magnitude fewer parameters, the right design and training can carry a model into a practically useful performance band. The day made the two diverging currents—scaling up and shrinking down, advancing simultaneously—even sharper.
Gartner Places Agentic AI at the "Peak of Inflated Expectations"
Gartner Japan published a Hype Cycle for the future digital workplace in Japan, placing agentic AI at the "Peak of Inflated Expectations." It also points out that while AI agents spread through organizations, the risk of "shadow AI"—operated at the discretion of frontline teams—continues to grow. A verdict that expectations have peaked should not be read as denying the technology's underlying capability; the natural interpretation is that it signals a phase where the "groundwork for moving into real-world operations" is about to begin in earnest.
Indeed, demand for operational guardrails has already become explicit. Server Works began offering a service that helps organizations draft guidelines for generative AI use in AWS environments, including a plan that can put rules in place in as little as one month. Efforts to support the shift from the "just try it out" stage to the "use it safely as an organization" stage are underway in the Japanese market as well.
Multimodal Models Take On Understanding and Generation in 3D Space
Research from major model developers is also advancing in domains that demand spatial understanding. The newly released Qwen-3D is a general-purpose vision-language model that leverages 3D geometric information as "a natural compression mechanism for visual streams." Using depth and camera pose as cues, it integrates observations from multiple viewpoints and multiple points in time into persistent representations aligned in world coordinates—a structure said to sidestep the difficulty of scaling to long contexts, where images and short video clips tend to hit a wall. It is drawing attention as a new approach to bringing spatial understanding into language models.
Tencent's Hunyuan3D-Buffalo 1.0 is a unified model that handles 3D understanding, text-to-3D generation, and instruction-driven editing within a single framework. After identifying the scarcity of geometrically consistent editing data as the barrier for unified 3D models, it attempts to ease this constraint through scalable training. Both efforts point in the same direction—from 2D image-centric modeling toward space—suggesting that the next major battleground for multimodal models is shifting to 3D.
Research Highlights: LLM "Conformity" and the Limits of Privacy Filters
Two studies have once again exposed limits in model behavior. The first identifies a phenomenon dubbed the "Artificial Hivemind" effect, in which LLMs converge on a narrow consensus even for open-ended questions, eroding diversity. Even under high-temperature sampling, similarity between responses reached 0.80–0.90, and the authors propose a mitigation combining meta-persona anchoring with sequential temperature scaling. For applications that require diverse perspectives, this is a property that cannot be overlooked.
The second is the first independent evaluation of OpenAI's "Privacy Filter," a filter for masking personal information. When the 1.5B bidirectional PII detector was validated across 22 languages, 5 domains, and 42 benchmarks, it outperformed existing methods for English-language PII detection, but trailed XLM-RoBERTa on multilingual NER across all 13 Indic and non-Latin-script languages. The results show that even a powerful filter becomes more fragile than expected when crossing language and domain boundaries.