According to Marketplace Pulse's 2026 Seller Index survey, 49% of Amazon sellers cite platform fees as their primary profit pressure, and 46% are concerned about advertising spend; however, only 24% are actively reducing their revenue dependence on Amazon. During the same period, fewer than 8,000 sellers contributed 50% of Amazon US third-party GMV, while the number of active sellers dropped from 584,000 in January 2025 to 500,000 in March 2026. Policy changes such as rising platform fees and extended payout cycles have exacerbated declining seller profits (47% reported year-over-year profit decreases). Meanwhile, Amazon's third-party segment revenue accounts for approximately one-third of the company's total revenue, while relying on sellers for 69% of GMV and 62% of order volume.
The updated Codex app (for macOS and Windows) adds computer use, in-app browsing, image generation, memory, and plugin features to accelerate developer workflows.
Hugging Face's official blog released a guide for training Sentence Transformers multimodal embedding and reranking models, using the fine-tuning of Qwen/Qwen3-VL-Embedding-2B for Visual Document Retrieval (VDR) as an example.
Hugging Face released a Skill for migrating transformers models to mlx-lm, along with a companion test suite, supporting automated environment setup and model downloading.
The Hugging Face team released the Ecom-RLVE framework, extending RLVE from single-turn reasoning questions to multi-turn, tool-augmented e-commerce dialogue agent tasks, offering 8 programmatically verifiable environments (product discovery.
以太坊基金会联合 Secureum、The Red Guild 和 Security Alliance（SEAL）于2024年底启动为期六个月的 ETH Rangers 计划，向17位独立安全研究者提供资助；结项成果包括追回或冻结超580万美元资金、报告785个漏洞与客户端缺陷、识别百余朝鲜IT从业者、交付7个开源工具库、开展80余场技术培训与36+次事件响应等。
Google DeepMind has released Gemini 3.1 Flash TTS, a new-generation text-to-speech model that supports over 70 languages and introduces audio tags for fine-grained control over tone, rhythm, and expression. It achieved an Elo score of 1211 on the Artificial Analysis TTS leaderboard and embeds SynthID watermarks by default.
IBM Research and Hugging Face released VAKRA, an executable agent evaluation framework based on real databases and documents. It covers four capabilities: API chaining, dashboard tool selection, multi-hop reasoning, and multi-source multi-hop with policy constraints. It includes over 8,000 local APIs and domain-aligned document sets. Evaluation uses a waterfall pipeline, supporting trajectory execution verification and LLM-assisted comparison. Error analysis indicates significant bottlenecks in model performance regarding tool selection, parameter filling, and policy adherence.
OpenAI updated the Agents SDK, adding native sandbox execution and model-native frameworks to help developers build secure, long-running agents across files and tools.
After HCompany released the Holo3 computer-use model on March 31, it launched the HoloTab Chrome extension, supporting zero-setup automation of web tasks; users can generate reusable routines by recording actions to automatically complete repetitive cross-website work.
The Tongyi Qianwen team has open-sourced Qwen3.6-35B-A3B, a sparse MoE model with 35 billion total parameters and only 3 billion activated parameters. Specializing in agentic coding, it surpasses Qwen3.5-35B-A3B and Qwen3.5-27B on multiple benchmarks, including SWE-bench Verified.
While evaluating its coding agents under the Maestro framework, AI21 found that the LLM judge acting as a reducer (using Claude Opus 4.5) significantly preferred superficially concise, minimal 'golden-like' patches over correct patches that actually passed tests; this bias persisted even on the contamination-controlled SWE-rebench dataset, indicating it stems from the model's inductive bias toward 'golden answer' stylistic features (such as minimality), rather than data leakage.
Google Research, in collaboration with New York University, launched the Vantage research experiment. It utilizes generative AI to build a multi-role simulated dialogue environment, where an Executive LLM dynamically guides conversations and an AI Evaluator analyzes dialogue text based on pedagogical rubrics. This enables scalable assessment of future-ready skills such as critical thinking, collaboration, and creative thinking. Experiments showed that AI scoring achieved consistency comparable to human experts, and English-language registration is now open via Google Labs.
Google DeepMind has released Gemini Robotics-ER 1.6, an upgraded reasoning-first model for embodied intelligence that significantly enhances spatial reasoning, multi-view understanding, task success detection, and gauge reading capabilities.
The Ethereum Foundation released Checkpoint #9, reporting on the progress of the Glamsterdam upgrade: the implementation of ePBS is more complex than expected, and non-headline features such as gas repricing also present technical difficulties; for Hegotá, the consensus-layer headline feature FOCIL (EIP-7805) has been selected, while Account Abstraction (EIP-8141) has been downgraded to CFI status; submissions for non-headline EIPs opened on April 9; the gas limit target is tentatively set at 60M, with higher values currently being tested; an EIP Champion handbook was also launched simultaneously.
ElevenLabs has launched three localized voice AI deployment methods: on-premise, on-device, and VPC. On-premise runs in customers' own data centers and confidential computing GPU environments; on-device supports offline inference, suitable for automotive and wearable devices; VPC is now available for deployment on AWS SageMaker and GCP Vertex.
Google Research released the paper ConvApparel, presenting a dataset in the apparel shopping domain containing over 4,000 human-AI multi-turn conversations (nearly 15,000 turns) to measure and narrow the realism gap of LLM user simulators.
Overworld released Waypoint-1.5 on Hugging Face, a new generation of real-time video world models that supports the real-time generation of interactive environments at 720p@60FPS on desktop graphics cards ranging from RTX 3090 to 5090.
The Sentence Transformers v5.4 update supports multimodal embeddings and reranking models for text, images, audio, and video, providing a unified API; it adds dozens of pretrained models such as Qwen/Qwen3-VL-Embedding-2B.
Google Research has released two agents: PaperVizAgent (formerly PaperBanana) and ScholarPeer. The former generates publication-quality academic figures based on method descriptions and captions, while the latter simulates the workflow of senior reviewers, combining literature retrieval with multi-dimensional verification to generate in-depth review reports. Both are experimental research prototypes and are not yet production-ready.
Safetensors announced its joining of the PyTorch Foundation, which is part of the Linux Foundation, becoming a foundation-hosted project alongside others such as DeepSpeed and vLLM; the format was created by Hugging Face.
Zhipu released the GLM-5.1 open-source large language model, claiming it can independently sustain work for over 8 hours in a single task, completing autonomous planning, execution, and self-evolution; it ranks third globally among models based on the average score across three code benchmarks: SWE-Bench Pro, Terminal-Bench 2.0, and NL2Repo.