Mistral opens Large 4 weights: 1T parameters, 52B active

Overview

9 stories in this issue. The first 3 are today's priorities.

Hot model updates

1. Top · Mistral opens Large 4 weights: about 1T parameters, 52B active
2. Top · Anthropic study: experts follow an agent's coin-toss preferences up to 86% of the time
3. Top · Sources say OpenAI's "Bel" is GPT-6.2 Betelgeuse

Global AI news

4. Tesla Robotaxi runs until 2 a.m. and adds Grok to Model Y
5. Epsilon builds a plain-English UI testing service on Amazon Nova Act
6. OpenAI-backed Thrive Holdings buys ADT for $15.3B enterprise value

Regional and early signals

7. Nvidia team opens a 4K video regeneration stack built on MiniMax H3
8. Baidu upgrades its search AI API with a ¥4.9 per 1,000 calls promo price
9. Doubao offers 8,800 free credits for 30 cumulative check-in days

Jiufeng graphic based on the sources cited in this issue.

Hot model updates

Mistral has published the Large 4 ("Le Chonk") weights on Hugging Face. It does well at coding but ranks behind Chinese open models overall.

Mistral released the Large 4 weights three weeks after the public preview. The model is nicknamed "Le Chonk." Mistral calls it the best open-weight model for coding and says it was trained entirely in Europe. The Decoder reports these specs and independent results:

• Parameters: about 1 trillion, 52B active (the Hugging Face repo is Mistral-Large-4-1T-A52B)
• License: modified MIT; it can run on your own hardware
• Modalities: text and image in, text out
• Artificial Analysis Intelligence Index: 26 points
• Output speed: 161 tokens per second
• Cyber safety: refuses 96.6% of harmful cyber requests

Limitations: "Best for coding" is Mistral's own claim. In the first independent test, Large 4 scores below Chinese open models such as DeepSeek V4.1 Flash and GLM-5.3 on the overall index, and well below the leading Anthropic and Google systems. The Decoder says its main strengths are speed and lower API prices. Self-hosting a 1T-parameter model also takes a lot of hardware.

Image source: Mistral.

Source: Original source Mistral · Hugging Face · The Decoder

On judgment calls with no right answer, experts mostly went with the AI agent's lean and believed the conclusion was their own.

Anthropic had 50 experts use an AI agent on four analytical tasks with no clearly correct answer. Each judgment call had two valid options, and a coin toss decided which one the agent preferred. Participants followed the agent's preference in up to 86% of cases, and most believed the final result was their own. Experience, expertise and trust in AI made no difference. Participants using AI also worked faster and made fewer serious errors.

Limitations: The study covers only 50 people and four tasks. The authors warn that if many teams use the same model, research could converge on a false consensus and AI safety checks could weaken. They suggest agents argue both sides or ask users for their own view first.

Source: Original source Anthropic · Paper (PDF) · The Decoder

Sources have given a full name and lineage for a rumored OpenAI model, but OpenAI has not confirmed it.

RuntimeWire sources say the reported "Bel" is short for GPT-6.2 Betelgeuse and that the model is based on the Dougtrio pre-train series.

Limitations: OpenAI has not publicly confirmed the model or shared its capabilities, pricing or availability. Developers have nothing public to plan around yet, so treat this as a rumor.

Source: Original source Post on X · RuntimeWire

Global AI news

Robotaxi now runs 20 hours a day and adds several ride features, including Grok on Model Y vehicles.

According to the official Tesla Robotaxi account on X on October 30th, service now runs from 6 a.m. to 2 a.m. Tesla had announced a midnight close in Austin on October 21st, so this adds two hours. The same update adds a rear camera feed during rides, trip sharing with a friend, a redesigned Live Activity, Calm Mode for the cabin screen and lighting, and Grok on Model Y vehicles.

Limitations: The details come from Tesla's own update post. The report gives no fleet size and does not say what Grok actually does during a ride.

Source: Original source Tesla Robotaxi on X · RuntimeWire

Publicis-owned Epsilon runs UI testing as an internal multi-tenant service where teams describe tests in plain English.

An AWS blog post describes Epsilon's setup. Epsilon is part of Publicis Groupe and handles hundreds of millions of consumer profiles and thousands of brand relationships, with frequent app changes. Its agentic platform team has 50 engineers. The team built managed UI testing on Amazon Nova Act and Amazon Bedrock AgentCore Browser. Tests are written in plain English, and the platform handles federated identity, pre-flight authorization, MCP integration and cross-account provisioning, so product teams can use it without extra setup.

Limitations: This is an AWS customer case study, written by AWS and the customer. AWS has also published sample code for Nova Act with AgentCore Browser that readers can try themselves.

Source: Original source AWS Machine Learning Blog · Sample code (GitHub)

A holding company built to rework traditional businesses with AI has bought home security company ADT.

Thrive Holdings announced it is buying ADT for $10 per share, about $7.6 billion in equity. ADT puts the deal's enterprise value at $15.3 billion. Thrive Holdings was spun out of Josh Kushner's VC firm Thrive Capital, and OpenAI is both a partner and an investor. The company holds businesses long-term and rebuilds them with AI. It currently holds more than 75 companies.

Limitations: The report only says Thrive plans to bring AI to home security. It does not name the AI products or models it will use, and gives no timeline.

Source: Single report TechCrunch

Regional and early signals

Open-H3-Regenerate-2K has open weights and can replace MiniMax's official 2K regeneration module, which is API-only.

Chinese-language source. The Efficient-Large-Model team, made up of Nvidia and MIT researchers, released the Open-H3-Regenerate-2K weights on Hugging Face, with inference code in the NVlabs/Sana repo. It turns low-resolution MiniMax H3 video drafts into 2560×1440 video, with native 4K output up to 3840×2160. Training keeps the H3 Transformer backbone and video VAE frozen and trains only a new refinement LoRA. Inference uses three denoising steps. The team tested a roughly 5-second clip on a single B200:

Limitations: The project team ran its own benchmarks. On the UniPercept perceptual score in RefinerBench-150, the open stack averages 60.0 and the official API 53.4. The official API is still slightly better on motion smoothness and dynamic degree.

Source: Single report QbitAI (Chinese)

At a Baidu app launch event, Baidu said the new search API's promotional price is below DeepSeek's search API.

Chinese-language source. At a Baidu app event in Beijing on October 30th, Baidu released an updated search AI API at a promotional price as low as ¥4.9 per 1,000 calls, which Baidu says is cheaper than DeepSeek's search API. The Wenxin assistant can deliver finished outputs for complex questions and covers more than 170 scenarios. The AI lens can recognize and track multiple objects and hold real-time conversations in 9 Chinese dialects. Interactive GUI scenarios on the results page have grown fivefold, drawing on more than 5 billion data resources, including government, public-service and central bank data.

Limitations: ¥4.9 is a promotional price. The report does not say how long the promotion lasts or give a price outside it. The claim that search has achieved "recursive self-evolution" is Baidu's own, with no public evaluation or technical details.

Source: Single report Leiphone (Chinese)

Every user earns credits for daily check-ins, and new desktop users also get a 30-day free trial plan.

Chinese-language source. New users of the Doubao desktop app or Doubao Work desktop app get a 30-day free trial on first login. All users get 200 credits per daily check-in, doubled on Mondays. Reaching 500, 2,000 and 5,000 credits earns bonuses of 200, 500 and 1,300, for 8,800 credits after 30 cumulative check-in days. The subscription page shows these plan quotas per 7 days:

Limitations: Bonus credits expire after 30 days. For non-paying users they cannot be used for video generation or the Pro model.

Source: Single report IT Home (Chinese)