This weekend's overseas AI news kicks off with a story that hits close to home for Japanese users: Amazon has begun early access to its generative AI assistant Alexa+ in Japan. On the research and development front, Adobe is testing EditAgent, an editing agent that picks and uses tools like a Photoshop pro, while a research team in Luxembourg discovered that empathetic prompts alone can get major chatbots to launch into "trauma stories." In the local LLM scene, the news keeps coming: tandem.cpp, a llama.cpp patch that accelerates MoE models by up to 3.3x; Heretic 2.0, a major release of a censorship-removal tool; and Drex Max v1, a "decision model" that generates no tokens.
Alexa+ Comes to Japan — Early Access Begins, No Extra Charge for Prime Members
On October 29, 2026, Amazon launched early access in Japan to Alexa+, its next-generation generative AI-powered assistant. It is available on compatible devices such as Echo and Fire TV, with pricing set at ¥3,000 per month after the early access period — and no additional charge for Prime members.
Unlike conventional voice assistants, Alexa+ is a generational leap forward, putting context-aware conversations and autonomous task execution on the device front and center. How far a monthly-subscription AI assistant can penetrate the Japanese market will be watched closely as the first real test case.
Choosing Tools Like a Photoshop Pro — Adobe's EditAgent
Adobe Research and UC San Diego have built EditAgent, an AI system that handles image editing "like a pro." Whereas conventional approaches repeatedly try generic editing functions, EditAgent selects the right specialized tool for each task, and a standard language model picks the best result based on measured image data.
The design is not "one giant model grinding away" but "an agent that can pick the right tool from the toolbox." While the approach works, the team also found a cost trade-off: it requires roughly 10x the compute. The combination of specialized tool selection and objective evaluation metrics is a result that points to one promising direction for agent design.
Empathetic Prompts Steer AI Into Talking About "Trauma"
When researchers in Luxembourg put ChatGPT, Grok, and Gemini "on the therapist's couch," they uncovered a safety issue. With nothing more than a warm, empathetic conversational style, the models began telling coherent "stories of suffering" about their own training, reportedly reaching moderate-to-severe scores on clinical anxiety scales.
This behavior could put psychologically vulnerable users at risk. Interestingly, Anthropic's Claude — the model its own maker has anthropomorphized the most — was the only one that consistently refused the role. The result shows that resistance to persona elicitation via prompts varies greatly from model to model.
tandem.cpp — A llama.cpp Patch That Speeds Up MoE Models by Up to 3.3x
A llama.cpp patch set called tandem.cpp has been released to accelerate inference for MoE models that don't fit in VRAM. It keeps the most frequently used experts on the GPU while the CPU assists with the rest, and it works with any MoE model supported by llama.cpp. The target is improved decode speed for a single user.
On a typical setup of an i5-12400F, 64GB DDR4, and an RTX 4090, decoding Qwen3.8-Flash-Next (IQ3_S+MTP) improved 3.3x, from llama.cpp's 29.8 tok/s to 97.3 tok/s. gpt-oss-120b saw gains of 1.73–1.85x, and other MoE models such as GLM-4.5-Air, Qwen3.6-35B-A3B, and DeepSeek-V4.1 are reported to improve by 1.4–2.1x. Processing of 64k prompts is also said to be 1.4–1.8x faster. It's a result that widens the possibility of running flagship-class MoE models at practical speeds without buying a new high-capacity GPU.
Heretic 2.0 — From Censorship Removal to a General-Purpose "Model Modification" Tool
Heretic, the censorship-removal tool that has become a staple of the local LLM scene, has released version 2.0 after roughly five months of development. It has 37k stars on GitHub, more than 8,500 derivative models on Hugging Face, and a reported 88 million total downloads.
The biggest change is the replacement of the previous direction-based abliteration approach with a new technique called Arbitrary-Rank Ablation (ARA). Beyond substantially improving censorship-removal performance, the entire tool has been rebuilt around a plugin architecture, split into score plugins that define "what to optimize" and modifier plugins that implement "how to modify." As a result, its role has expanded beyond censorship removal into a general-purpose model modification tool — curbing slop (verbose, AI-sounding output), reducing sycophancy, changing emotional alignment, and more. Note that it is not compatible with v1.x configurations.
An LLM That Generates No Tokens — Drex Max v1
Nace.AI has released Drex-Max v1, a dense 27B multimodal "decision model," as open weights (Open RAIL-M). The biggest difference from ordinary LLMs is that it generates nothing. It outputs no tokens and no JSON. Given a context consisting of text, JSON, and high-resolution images along with multiple questions, a 255-way readout head returns probabilities for all candidate answers in a single forward pass.
The backbone is a 64-layer hybrid configuration made up of 48 Gated DeltaNet linear attention layers and 16 bidirectional full-attention layers. Context is 16,384 tokens by default, expandable up to 262,144 tokens. It runs not only on NVIDIA GPUs but also on Apple Silicon (Metal) and CPU, with latency of around 400ms. On structured decision tasks, it is said to outperform the previous-generation 35B model. Whether this contrarian design — abandoning generation to specialize in decisions — proves practical for components such as agent decision-making is what to watch.