Mistral rolls out Le Chonk

Mistral put a trillion-parameter open-weight model into public preview as companies ask how much control they need over AI.

Today’s issue covers Le Chonk, customer-agent access, and SAP’s agent push inside core business systems.

Mistral has launched Mistral Large 4 into public preview, and its pitch is about control as much as raw model power.

Most buyers have four choices: buy a finished AI app, connect vendor-hosted AI like ChatGPT, Claude or Gemini to company data, privately run or adapt an open-weight model, or build a model from scratch.

Mistral’s new model, nicknamed “Le Chonk”, is aimed at the third choice: more control than a vendor-hosted API like ChatGPT, Claude or Gemini, without starting from zero.

The company says ML4 has 1 trillion parameters, with 49 billion active at a time. In plain English, it is a very large model that only switches on part of itself for each request, which helps control cost.

Mistral says the model is strong in cyber, finance, law, coding and work that uses images as well as text.

Two deployment details matter:

Mistral also says ML4 was trained on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres. The pitch is partly about model performance, and partly about where the model runs, who controls it, and which legal system surrounds the data.

Founders context: Do not buy “open” as a slogan. Ask what actually changes: where the model runs, who can see the data, who updates it, how much monitoring you inherit, and whether the extra control is worth the added work.

Sierra and Meta are trying to help companies tell the difference between a real customer’s authorised agent and an ordinary bot hitting a checkout, booking flow or support line.

They have introduced Personal Agent Protocol, an open standard being developed with Genesys, Instinct, Rocket, Shopify, Stripe and Walmart.

The proposal is a front door for authorised agents. A person gives an agent permission to act, then a company decides what that agent can do through its website, APIs or its own customer-service agent.

That matters because personal agents are moving from demos into jobs such as scheduling appointments, booking travel and shopping for insurance. Without a cleaner connection, many will behave like people using a browser: loading pages, filling forms, calling support or opening web chat.

Sierra says a direct connection should make that faster and safer.

The business question is not whether every customer will use an agent next month. It is whether companies can recognise the difference between a human, a bot, and an authorised agent acting for a real customer.

A usable setup needs a few basics:

This is how agent commerce starts looking less like web scraping and more like account access with rules.

Founders context: If customers may send agents into your checkout, booking flow or support line, map the first three tasks you would allow. Then decide what permission, spending limit and audit trail each task needs.

SAP used its Connect event to show agents moving into the least glamorous parts of company work: procurement, collections, billing and supply-chain planning.

That is the practical path for business AI. The agent is not floating around the company with a vague mandate. It sits inside core software, reads trusted data, suggests the next step and leaves a person close enough to approve it.

SAP named several customer examples.

Novartis has co-developed and recently launched a pilot of SAP’s Joule Sourcing Assistant. Morgan Foods says it is preparing for agents to connect demand signals, inventory levels and supplier status faster than people can manage manually.

Nestlé is using its SAP S/4HANA rollout across 140 countries as the base for an Accounts Receivable Assistant that can prioritise collections work by risk and business context.

PwC is using SAP’s Billing Assistant to catch possible billing issues earlier.

SAP also says Joule Work is live with 110,000 SAP employees and has produced 20% productivity gains across finance, HR and procurement.

Useful agents are starting in structured systems with clear fields, permissions, approvals and audit trails.

Founders context: Start with the business system that already holds the work. Agents are easier to judge when they act on known data, follow existing approval paths, and leave a record your team can inspect.

🛡 Anthropic expanded cyber access for vetted defenders. Its Cyber Verification Program now has three tiers for approved security teams that need stronger models with fewer cyber refusals.

📱 Google brought multimodal search to the edge with EmbeddingGemma 2. The 740M-parameter open-weight model can search across text, images, video frames and audio locally on devices.

🏔 Sierra Summit added new customer-service models and Tandem Voice. Sierra says Curie handles the conversation loop, while Fleming detects when a caller is another agent.

🧮 TechCrunch put Mistral’s model into the open-model rivalry. Its write-up focuses on how ML4 is aimed at both closed American models and open Chinese rivals.