Gate News message, April 23 — DeepSeek has open-sourced TileKernels under the MIT license, a GPU kernel library written in TileLang for large language model training and inference. TileLang is a domain-specific language developed by the tile-ai team for expressing high-performance GPU kernels in

Gate News message, April 23 — Tencent unveiled and open-sourced Hunyuan Hy3 preview, a hybrid mixture-of-experts language model featuring fast and slow thinking fusion. The model comprises 295 billion total parameters with 21 billion active parameters, supporting a maximum context length of 256K

Gate News message, April 23 — Google researchers, including He Kaiming and Xie Saining, published a paper introducing Vision Banana, a general-purpose vision understanding model created through lightweight instruction fine-tuning of the company's Nano Banana Pro (Gemini 3 Pro Image) image

Gate News message, April 23 — ByteDance's Seed team released Seed3D 2.0, a text-to-3D model that generates textured 3D assets from a single image. The upgrade focuses on geometric precision and material realism, with the API now available on Volcano Ark. Geometric generation employs a

Perplexity uses SFT followed by RL with Qwen3.5 models, leveraging a multi-hop QA dataset and rubric checks to boost search accuracy and efficiency, achieving best-in-class FRAMES performance. Abstract: Perplexity's post-training workflow for web-search agents combines supervised fine-tuning (SFT) to enforce instruction-following and language consistency with online reinforcement learning (RL) via the GRPO algorithm. The RL stage uses a proprietary multi-hop verifiable QA dataset and rubric-based conversational data to prevent SFT drift, with reward gating and within-group efficiency penalties. Evaluation shows Qwen3.5-397B-SFT-RL achieving top FRAMES performance, 57.3% accuracy with a single tool call and 73.9% with four calls at $0.02 per query, outperforming GPT-5.4 and Claude Sonnet 4.6 on these metrics. Pricing is API-based and excludes caching.

Abstract: OpenAI's Privacy Filter is an open-source, locally executable model that detects and redacts PII in text. It supports large contexts, identifies many PII categories, and is intended for privacy-preserving workflows such as data preparation, indexing, logging, and moderation. OpenAI's Privacy Filter is a locally run, open-source model (128k-token context) that detects and redacts PII in text, covering contact, financial, and credential data for privacy workflows.

Gate News message, April 22 — Princeton PhD student Yifan Zhang disclosed complete technical specifications for DeepSeek V4 on X, following a preview on April 19. V4 features 1.6 trillion total parameters and a lightweight variant, V4-Lite, with 285 billion parameters. The model employs DSA2

Kimi K2.6 tops OpenRouter leaderboard, outperforming Claude, GPT, and open-source rivals, signaling domestic AI progress and narrowing the gap with global leaders. Abstract: Kimi.ai announced that its latest model, Kimi K2.6, ranked first on the OpenRouter programming-ability leaderboard, leading developer evaluations. Benchmarks indicate K2.6 delivers superior performance across programming tasks relative to Claude, GPT-series, and other open-source models, highlighting gains in code generation and development-task handling and signaling progress for domestic AI toward international leaders.

Altman on Core Memory recounts failed OpenAI governance talks with Elon Musk: stages of compromise toward a for-profit model, Musk's demands for majority stake and CEO control, Altman opposing absolute power, with trial looming. Abstract: Sam Altman details, on Core Memory, failed negotiations with Elon Musk over OpenAI governance, outlining moves toward a for-profit model, Musk's demands for majority stake and CEO authority, and Altman's rejection of absolute control; the litigation looming with a trial set for April 27.

Gate News message, April 22 — Google Research released ReasoningBank, an agent memory framework that enables large language model-driven agents to continuously learn after deployment. The framework extracts universal reasoning strategies from both successful and failed task experiences, storing