Scoutari

AI builder 的決策情報台。

瀏覽

  • 今日情報
  • Skill 榜單
  • AI 週報
  • 關於與編輯方針
  • RSS

主題

  • 模型發布
  • API 與價格
  • MCP 與 Skills
  • AI Coding
  • 中國模型
  • 開源模型
  • Agent
  • Infra / 成本
  • 值得試用
© 2026 Scoutari · scoutari.com摘要由 AI 輔助生成,並始終附上原文連結。
情報收藏週報設定
SScoutariAI Builder Intel · decision desk
Skill RadarSavedPersonalizeSubmit sourceLog inSign up
Sign up
Skill RadarSavedPersonalizeSubmit sourceLog in

Chinese models

Qwen, DeepSeek, Kimi, Zhipu GLM, MiniMax, and other Chinese frontier labs.

Model releasesAPI & pricingMCP & SkillsAI codingOpen sourceAgentsInfra & costMultimodal & imageAI videoVoice AIAI hardwareRobotics & embodiedAI safetyResearch & papersEnterprise AIAI fundingPolicy & regulationWorth trying
  • Mon, July 6Open Source

    MaxKB: Open-Source Enterprise Agent Platform Trending on GitHub

    MaxKB is an open-source enterprise agent platform gaining significant traction on GitHub.

  • Mon, July 6Open Source

    DeepSeek-Reasonix: Terminal-native AI coding agent with prefix-cache stability

    esengine/DeepSeek-Reasonix is a DeepSeek-native terminal AI coding agent designed for prefix-cache stability.

  • Mon, July 6Tools

    Open Design: Local-First Desktop App Turns Code Agents into Design Engines

    Open Design is an open-source, local-first desktop alternative to Claude Design, enabling code agents to generate designs and export to HTML/PDF/PPTX/MP4.

  • Fri, July 3Model/API

    DeepSeek Peak Pricing, Goldman Sachs Says China's LLM Competition Turns Rational

    DeepSeek raises API prices during peak hours, signaling a shift from subsidy-driven competition to sustainable pricing in China's AI model market.

  • Thu, July 2Model/API

    Meituan Limits Doubao LLM Usage Internally

    Meituan reportedly restricts internal use of Doubao large language model.

  • Thu, July 2Tools

    Kimi K2.5 Code Officially Integrated into GitHub Copilot

    Kimi K2.5 Code is now generally available in GitHub Copilot.

  • Tue, June 30Open Source

    DeepReinforce Launches Ornith-1.0: Self-Constructed LLM for Agentic Coding

    DeepReinforce releases its first open-source model, Ornith-1.0, based on Gemma 4 and Qwen 3.5, achieving state-of-the-art results across coding benchmarks.

  • Sun, June 28Model/API

    Liquid AI Launches Smallest Model LFM2.5-230M with Multi-Framework Support

    Liquid AI releases LFM2.5-230M, supporting llama.cpp, MLX, vLLM, SGLang, and ONNX, achieving 213 tok/s on Galaxy S25 Ultra.

  • Sat, June 27Model/API

    ByteDance Diffusion LM iLLaDA Matches Qwen2.5

    ByteDance and Renmin University unveil 8B parameter diffusion language model iLLaDA, matching Qwen2.5 base performance.

  • Fri, June 26Model/API

    AI Startup Lindy Switches to Deepseek, CEO Says Saves Millions

    AI startup Lindy fully switched from Claude to Deepseek, citing AI costs exceeding personnel costs as key to survival.

  • Fri, June 26Open Source

    DeepReinforce Open-Sources Ornith-1.0 Coding Models with Self-Learning RL Scaffolds

    DeepReinforce releases Ornith-1.0, an open-source coding model family based on Gemma 4 and Qwen 3.5, featuring self-learned RL scaffolds, with the 397B flagship scoring 82.4 on SWE-Bench Verified.

  • Thu, June 25Model/API

    Snowflake CEO Finds GLM-5.2 Costs One-Fifth of Opus 4.7, Nears Performance

    GLM-5.2 achieves near Claude Opus 4.7 performance on Snowflake benchmark at one-fifth the cost.

  • Wed, June 24Research

    DFlash Speculative Decoding: Parallel Token Blocks Boost Throughput Up to 15x on NVIDIA Blackwell

    UC San Diego's DFlash replaces autoregressive draft generation with a lightweight block diffusion model, generating entire token blocks in a single forward pass for accelerated speculative decoding.

  • Tue, June 23Open Source

    Prime Intellect Releases prime-rl 0.6.0 for Trillion-Parameter MoE Training

    Prime Intellect released prime-rl 0.6.0, an open-source framework for async reinforcement learning on trillion-parameter mixture-of-experts models.

  • Sat, June 20Model/API

    VibeThinker-3B: 3B Dense Reasoning Model via Spectrum-to-Signal Post-Training

    VibeThinker-3B is a 3B-parameter MIT-licensed reasoning model matching DeepSeek V3.2 and Kimi K2.5 on verifiable benchmarks.

  • Fri, June 19Model/API

    GLM-5.2 passes vibe check, open-source models enter frontier

    GLM-5.2 has passed the collective vibe check, marking a milestone for open-source models.

  • Thu, June 18Model/API

    China Z.ai Lab Releases GLM-5.2, the Strongest Open-Source Text LLM

    Z.ai Lab open-sourced GLM-5.2, a 753B parameter text-only LLM, on June 16.

  • Wed, June 17Model/API

    MiniMax Unveils MSA: Dual-Branch Sparse Attention for MoE

    MiniMax released MSA, a dual-branch block sparse attention mechanism that reduces compute costs.

  • Wed, June 17Model/API

    Microsoft Copilot Cowork Shifts to Usage-Based Pricing, May Adopt DeepSeek

    Microsoft considers using a fine-tuned DeepSeek V4 as a low-cost model option for Copilot Cowork and shifts to usage-based pricing.

  • Wed, June 17Model/API

    Qwen-RobotSuite: Three Embodied AI Models for Vision-Language Manipulation, Video Modeling, and Navigation

    Qwen-RobotSuite includes RobotManip, RobotWorld, and RobotNav, targeting manipulation, video world modeling, and navigation.