An open directory of AI tools — make AI pay off

OpenAI will acquire Pierre Computer Company, expanding Codex’s tools for building, understanding, and reviewing code.

Discover how Heimstaden uses OpenAI’s ChatGPT Enterprise and Codex to transform property management, cut costs, and empower employees across Europe.

Today, we are announcing the Special Interest Group (SIG) for Ray on hardware accelerators. When Ray was created in UC Berkeley over 10 years ago, the computing landscape looked v…

On 2 November Meituan's LongCat released LongCat-2.5, a natively multimodal model for long-horizon agent tasks: 1.6 trillion total parameters with about 48 billion activated per token, strengthening native visual understanding and agentic coding over 2.0. Visual signals are encoded as multi-level discrete tokens so the backbone handles images and text uniformly; training adds long-horizon trajectories and large-scale long-horizon RL, turning real agent failures into reproducible training tasks. The stack is tuned for long-context training and W4A4 inference on domestic accelerators. LongCat says the model is strong on visual coding and supports creative design and 3D work; it is live at longcat.chat with an API on its platform page.

On 31 October Alibaba and Apple announced that Chinese users can call Qwen directly from Siri, Writing Tools and Visual Intelligence for text and image understanding and content generation. Testing opens with the latest iOS developer beta on 4 November, with general availability for Chinese users in the iOS 27.3 release in December, for iPhones bought in China that support Apple Intelligence or devices whose location and Apple Account region are both set to China. Users choose whether to enable it and what to share; Alibaba says requests and responses are not used for training.

On the evening of 31 October the DeepSeek Harness team published a technical post titled Super-agent design exploration: toward a more autonomous agent system, and open-sourced the prototype. The design has two layers: a Router that runs continuously as a global coordinator, taking messages from group chats, GitHub and email and dispatching tasks, and Workers that keep their own context and work in parallel on reproduction, code review and feature work, calling Subagents when needed. Event-driven triggers and a two-stage filter let it spot defects in group chats without being mentioned, then reproduce and attempt fixes in the background. The team says an internal instance already runs in daily R&D, and frames the post as engineering practice rather than a product announcement.

On 30 October Moonshot's 3D Make went live on Kimi's web and desktop apps, powered by K3.1 Pro and currently free. Given a sentence, a sketch or a reference image, the model infers shape, proportion and spatial relations, then writes Blender or FreeCAD scripts and runs them in the cloud — effectively driving professional modelling software remotely. Output can be downloaded or previewed in the browser, revised in natural language with instant recomputation, and exported in several formats with scene assembly. The feature first appeared with K3.1 Pro on 24 October. Printability still needs a human check: late-October research from MIT and partners rebuilt 120 popular Thingiverse models with AI and found 94 with two or more defect classes, 2.4 on average — unopened holes, thin walls, stray blobs.

We are excited to announce the release of PyTorch® 2.15 (release notes)! The PyTorch 2.15 release features the following changes: Rewritten NCCL backend becomes the default, with…

On 29 October Doubao Work opened third-party model access to enterprise customers, starting with Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, for Team and Enterprise subscription tiers. The feature is on by default: companies do not need a separate contract with each model vendor or an extra application, and usage is deducted from their existing AI credits.

Genie 4: our new world model unlocking unprecedented realism in real-time simulation.

On 28 October Tencent's WorkBuddy added remote control: one tap in the phone app connects to the computer, showing the desktop live and letting you operate it directly. The desktop appears as a small floating window above the chat, with the input box still usable so you can correct a task the moment it drifts; switching to full screen in landscape makes taps land on the same spot on the PC. Tencent says only one authenticated phone may connect to a computer at a time, the stream and input run over an encrypted channel, keystrokes are not recorded and the remote screen is not stored, and remote sessions consume data but no tokens.

On the morning of 28 October Doubao Work upgraded its desktop control: clicks are more precise and faster, and a full session can be recorded into a skill — the user demonstrates once and narrates the rules, and Doubao turns the clicks, typing and window contents into a personal skill that can be reused for similar tasks. A setting lets it keep working while the Mac is locked, with instructions and results available from the phone.

Forgotten projects can stay published, connected to data, and unowned. This Halloween, hunt them down in our free game, and see how Enterprise admins can clean up abandoned projec…

On 17 October DeepSeek moved V4.1-Flash to the V4.1-Flash-1017 build across web, app and API. The company says self-feedback reinforcement learning strengthened building whole projects from scratch — demos include a 3D model of the Forbidden City, a browser racing game and a Godot remake of Portal 2 — pitched as "wish-driven programming". Two in-house APIs, Search and Fetch, opened at ¥5 per thousand calls, free until 1 November.

On 15 October StepFun put the Step 5 Preview weights on Hugging Face and ModelScope for free download: 600B total parameters, 27B activated per token and a 1M-token context, aimed at long-horizon agent work in coding, software engineering and finance. At its 20 September debut it scored 44 on the Artificial Analysis Intelligence Index, placing in the global open-model top three, and the company puts per-task cost at roughly one-eighth of Claude Opus 5. Token volume on OpenRouter ranked second worldwide on launch day, and it also landed inside Tencent WorkBuddy and ByteDance TRAE for people who will not self-host.

On 14 October the 30-day comment period closed for Doubao's screen-automation declaration protocol (SAEP), and the accompanying opt-out list named 155 apps: 66 from Tencent (WeChat, QQ, Tencent Meeting, Honor of Kings), 23 from Alibaba (Taobao, Alipay, DingTalk, Amap), 14 from Kuaishou, 11 from Meituan, plus DeepSeek, Kimi, Pinduoduo, Xiaohongshu, WPS, several banks and all three carriers, across the Nubia M153 and NaviX Ultra. What is switched off is not installation but the assistant's ability to read the screen and simulate taps. StepFun's agent phone went on sale the same day with 100-plus apps wired into its Amoo — the first barrier for phone AI is no longer technical.

The 35B-A3B build is a mixture-of-experts model that activates about 3B parameters per token and keeps the rest in flash, prefetching them with its own lookahead expert scheduling: about 13 tokens/s in roughly 3 GB of memory on an iPhone 17 Pro Max, and 17+ tokens/s on an M4 iPad Pro. A 1B version targets lighter devices. Apache 2.0, commercially usable, with day-one support across iOS, Android, HarmonyOS and major chip platforms.

Caijing, citing several people involved in the deal: the raise grows from a planned ¥50B to about ¥100B, split roughly evenly between new and existing backers, with the valuation unchanged by the doubling. CATL is set to add about ¥10B, while Tencent and IDG return; industrial capital leads the new entrants. A STAR Market listing is being prepared in parallel.

Image generation and editing are now callable by developers, with Turbo weights open-sourced alongside. This generation builds on a 7B visual-generation architecture, cutting denoising to 8 steps while keeping 2K output, and accepts up to 10 reference images with region editing via selection, brushing or masks.

On 6 October the Amazon Web Services model platform added Zhipu GLM-5.3, sharing revenue based on call volume. Earlier, on 30 September, US AI infrastructure company Baseten said enterprise users could call Moonshot and Zhipu models from OpenAI's Codex coding tool, with fees drawn on existing OpenAI purchase commitments and no new vendor process — the first time Chinese open models entered OpenAI's enterprise billing channel.