HAI Daily

Kimi K3.1 Pro Opens Its Weights, Tencent Ships Hy4 and OpenAI Buys Pierre

Moonshot releases the weights of Kimi K3.1 Pro, Tencent makes Hy4 generally available, Chinese tech firms sign an agent self-regulation pact, and OpenAI buys Git tooling maker Pierre for Codex.

By H.AI Tools Team4 stories
Products in this issue: Kimi, Tencent Yuanbao, DeepSeek, Qwen Chat, Cline, Cursor, GitHub Copilot, Devin
In this issue

The short version

  • Kimi K3.1 Pro's weights are now downloadable. Moonshot reports 65.7% on Terminal-Bench 4.0, close to the top closed models, but running it yourself needs data-centre hardware.
  • Tencent's Hy4 is generally available with image understanding and computer use, at the same API price as the preview.
  • Twenty Chinese organisations, led by CAICT, signed a self-regulation pact for AI agents covering registration, interoperability, security and user rights.
  • OpenAI agreed to buy Pierre, whose Git tools already power parts of Codex. Its open-source projects will stay open.

November 5 was a day for coding and agents. Two Chinese labs made strong models more widely available, one as downloadable weights and one as a finished product. Chinese companies agreed on common rules for agents, and OpenAI bought the team behind the Git tools inside Codex. The sections below explain each story and how to use it in your own development work.

01

Kimi K3.1 Pro: open weights close to the frontier

What Moonshot released

Moonshot AI (月之暗面) published the weights of Kimi K3.1 Pro on Hugging Face on November 5, about a day later than it had promised. K3.1 Pro is a large mixture-of-experts model with a context window of about one million tokens, built on Kimi K3 with further reinforcement learning for agent tasks.

On Terminal-Bench 4.0, which tests whether a model can complete real tasks in a command-line environment, Moonshot reports 65.7%. That is ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%) on the company's own chart. Independent evaluator Artificial Analysis measured about 67%. One report says the model ranks first among open-weight models on the Agent Arena leaderboard; we have seen that ranking in only one source. Moonshot also removed cyber-attack training data from the open release, which it says lowers the model's exploit success on its internal test.

The weights are very large. Running the full model needs a multi-GPU server, so most people will use it through Kimi's own apps, its API or a cloud provider.

Trying an open model in your coding workflow

Open weights matter even if you never download them. More providers can host the model, prices tend to fall, and companies with strict data rules can run it inside their own cloud.

The practical test is simple. Pick five real tasks from your backlog: a bug fix, a small feature, a test-writing job, a refactor and a script. Run each in an agent such as Cline, which lets you choose the model provider, using K3.1 Pro and then your current model. Record whether the change passed your tests, how many retries it needed and the cost. Open models from DeepSeek (深度求索) and Qwen (通义千问) belong in the same comparison. Our Cline vs Cursor vs Roo Code comparison explains the agents themselves.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

Tencent Hy4: generally available with computer use

What Tencent announced

Tencent (腾讯) released the general-availability version of Hy4 on November 4. It is a 770-billion-parameter model and the first in Tencent's Hunyuan (混元) family to combine image understanding with the ability to operate a computer. Tencent says it outperforms open models such as DeepSeek-V4.1-Flash on several coding and computer-use tests, while still trailing the leading closed models. The API price is unchanged from the preview, at about $0.83 per million input tokens and $2.50 per million output tokens, and Tencent says a cheaper Flash version is coming. Our Special Edition for November 1 to 7 puts Hy4 alongside other agents that gained new access that week.

Tencent says the model is available across its products, including Tencent Yuanbao (元宝), and reports a lower rate of factual errors in financial analysis than the preview. These are company figures.

Where Hy4 fits

Hy4's price puts it in the budget tier, while its computer-use ability targets office tasks such as filling forms, moving data between applications and checking reports. If you already work in Tencent's ecosystem, test it on a repetitive desktop task in a test account. Watch every step on the first runs, and keep it away from payments, deletions and messages to customers until it has proved reliable.

03

A self-regulation pact for AI agents in China

What was signed

At the World Internet Conference in Wuzhen on November 4, twenty organisations published a self-regulation convention for AI agents. The China Academy of Information and Communications Technology (CAICT) led the drafting with companies including Tencent, Alibaba (阿里巴巴), Ant Group (蚂蚁集团) and Douyin (抖音). The signatories include Huawei (华为), Xiaomi (小米), Honor (荣耀), OPPO and Baidu (百度); published lists vary slightly between reports. The convention sets out eight commitments: building shared infrastructure for agents to find and connect with each other, including registration and discovery standards; interoperability; security protection; governance; user rights; helping industries adopt agents; and lowering costs for smaller companies. Signatories also pledged to work against misuse of agents for fraud and grey-market activity.

What it means for teams deploying agents

A voluntary pact is not law, but it signals what large platforms will expect from partners. If you deploy agents that interact with Chinese platforms, expect requests to register them, describe what they can do and show how you handle user data.

Prepare a one-page record for each agent you run: its purpose, the data it reads, the actions it can take, who approves those actions and how a user can complain. That record will help with platform reviews and with internal security checks anywhere. The AI Security category lists tools that help monitor model and agent behaviour.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

04

OpenAI buys Pierre to strengthen Codex

What OpenAI announced

OpenAI agreed to acquire Pierre Computer Company on November 4. Pierre makes Code Storage, a Git hosting service, and the open-source Trees and Diffs libraries, which already power the file browser and code-review views in Codex. OpenAI says Pierre's open-source projects will stay open, existing Code Storage customers can keep using the product, and its features will gradually move into Codex. The deal still needs regulatory approval.

Git habits for working with coding agents

The acquisition reflects a real problem: Git was designed for people, and agents produce far more branches, commits and changes. Whatever tools you use, a few habits keep agent work reviewable.

Give each agent task its own branch. Ask for small changes with a clear description, and reject large mixed diffs. Require tests to pass before review. Read the diff yourself, because a summary written by the agent is not a review. Tools such as GitHub Copilot, Cursor and Devin all work within these limits. Our Devin vs GitHub Copilot vs Augment Code guide compares how much autonomy each one takes.

What to try next

Test the new models on your own tasks

Run five real tasks through Kimi K3.1 Pro and your current model, and keep the one that passes more tests per dollar. Try Hy4 on a repetitive desktop task in a test account. Write a one-page record for each agent you run, and keep every agent change on its own small branch.

Explore Code Generation, Chatbots and AI Security 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.

Go deeper