In this issue
The short version
- DeepSeek's V4.1-Flash-1017 update finishes tasks with shorter outputs, and its new Search and Fetch APIs let agents look things up on the web. Billing starts on November 1.
- Tencent's Hy Memory 2.0 gives agents chat, work and code memory, with a small specialised model and a local app meant to share memory across AI tools.
- Sugon 8000, China's first fully domestic 100,000-card AI supercluster, runs both scientific simulation and large-model training on one system.
- Anthropic's $63 billion chip financing, arranged through Broadcom, Wall Street banks and Blackstone, shows AI computing is now funded like large infrastructure.
October 19 brought practical upgrades for people building agents, a better model with web search from DeepSeek (深度求索) and an open memory system from Tencent (腾讯), along with two stories about the hardware and money underneath. Each section explains what changed and how to use it.
01
DeepSeek V4.1-Flash-1017 and the new Search API
What DeepSeek released
DeepSeek released V4.1-Flash-1017 on October 17, an update to its V4.1 Flash model that keeps the same API name and price. DeepSeek says it improves benchmark results while shortening task time and output length, so the same work costs less. Reinforcement learning based on self-feedback improved its ability to build complete projects from scratch, including games and 3D modelling, and the open weights are on Hugging Face under the MIT licence.
DeepSeek also opened its own search service to developers through two new endpoints: Search, which returns web results, and Fetch, which retrieves a page's content. Chinese reports give the price as ¥5 per 1,000 calls for each, while English sources list $0.75 per 1,000; check DeepSeek's price page. Calls made through tools such as DeepSeek Harness had not been billed, and per-call billing starts on November 1. Both endpoints are limited to five requests per second per account.
Adding web search to your agents without surprises
A search API lets an agent check current facts instead of relying on its training data. It also adds cost and new failure modes: wrong pages, outdated pages and pages written to mislead AI. Three rules help. Ask the agent to quote the source URL for every fact it uses. Limit searches per task, for example five, so a confused agent cannot run up a bill. And treat fetched pages as untrusted input: an agent should never follow instructions written on a web page.
If you use agents through DeepSeek, check your usage before November 1 so the new charges do not come as a surprise. Answer engines such as Perplexity are the simpler choice when you want cited answers without building an agent. Our DeepSeek vs Kimi comparison covers two leading Chinese assistants.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
Tencent Hy Memory 2.0: long-term memory for agents
What Tencent released
Tencent's Hunyuan (混元) team released and open-sourced Hy Memory 2.0 on October 16, a memory system for AI agents in three layers. The first is a specialised model, Hy-Memory-30A3B, with about 3 billion active parameters. The second is a framework that stores, updates and develops three kinds of memory: chat, work and code. The third is MemHub, a local app meant to synchronise memory across about ten AI tools, including Cursor and Codex. Tencent says the framework ranks first among open-source memory frameworks on four benchmarks and that the specialised model outperformed Claude Opus 5 on its tests; these are Tencent's own figures. It also reports large accuracy gains when internal assistants used the system instead of the full conversation history.
One report notes that MemHub's own changelog said it had not yet been publicly released at the time, so check the official repository for availability.
Giving your agents a useful memory
Agents forget everything between sessions unless you give them memory, and a bad memory is worse than none: outdated facts get repeated with confidence. Start simply. Keep a short, human-readable file for each project with decisions, conventions, open questions and things the agent got wrong before. Coding tools such as Cursor and Cline, and assistants such as Claude, can load such files at the start of a session.
Review the file every week and delete what is no longer true. Only move to an automated memory system once you know what is worth remembering. Keep secrets and personal data out of any memory store. Our Cursor vs GitHub Copilot comparison covers how each handles project context.
03
Sugon 8000: China's first fully domestic 100,000-card supercluster
What was shown
At the 2026 National High Performance Computing Conference in Shenzhen, Sugon (中科曙光) showed the Sugon 8000, described as China's first fully domestic AI supercluster with 100,000 accelerator cards linked by high-speed interconnects. It follows a "native fusion" of supercomputing and AI computing: one system handles high-precision scientific calculation as well as large-model training and inference. Sugon says it can support projects such as a digital twin of the brain or simulations of the universe's evolution. The system was completed in July, connected to China's national supercomputing internet and reportedly ran at full load in its first week.
Why combining simulation and AI matters
Simulation works forward from physical laws, while AI learns patterns from data. Combining them lets researchers check AI predictions against physics, which reduces errors where data is scarce. For companies, the practical effect is more domestic capacity for engineering simulation and model training through national computing services. If your work involves simulation, such as materials, fluids or structures, ask your computing provider whether combined simulation-and-AI workflows are available rather than running them on separate systems.
04
Anthropic's $63 billion chip financing
Reports in mid-October said Broadcom was arranging more than $60 billion of debt, in senior and subordinated layers, to fund AI chips for Anthropic, with Blackstone providing the riskier subordinated part. The deal was announced at $63 billion on October 26, with Bank of America, Citi and Morgan Stanley arranging the senior portion and Broadcom supporting the lenders. Early reports said the riskier layer would pay more than Anthropic's earlier chip financing, but we could not confirm the final pricing. Either way, AI computing is now funded much like large infrastructure projects, with lenders and chip makers sharing the risk.
What to try next
Put one change into practice
Cap the number of web searches per agent task and require a source link for every fact. Start a plain project memory file for your agents and review it weekly. When you read about AI capacity, notice who carries the risk as well as how much is raised.
Explore Chatbots, Code Generation and Search Engines to build your shortlist.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.