The Intake — Tuesday, October 6, 2026

If you've been watching for open-weight frontier performance without API dependency, Mistral posted October 27 as the date.

Mistral released Large 4 — nicknamed "Le Chonk" — on October 6, 2026. It has 1 trillion total parameters with 49 billion active per forward pass. The architecture is mixture-of-experts. Mistral says training ran on Nvidia Grace Blackwell GPUs in their European data centers. The run took approximately two months.

Mistral reports 82% on the Artificial Analysis Cyber Index — the highest of any model evaluated at launch. It reports 62% on software engineering benchmarks. API preview is live now on Mistral Studio. Open weights are planned for October 27. The release will carry a custom Mistral license.

Mistral VP of Science Pierre Stock said the release was structured with "trusted partners and governments" to permit defensive security use while limiting offensive applications. The license text has not been published. If you're planning self-hosted or local deployment and the terms matter for your use case, what the license permits is not yet verifiable.

If your team's compliance requirements prohibit source code from transiting external systems, IBM Bob's self-hosted deployment option now exists.

IBM announced on October 1, 2026 that IBM Bob now supports customer-managed deployments. Supported configurations include on-premises, private cloud, sovereign cloud, and air-gapped environments. IBM Bob is IBM's full software lifecycle agent, covering planning, coding, testing, deployment, and legacy modernization.

Supported on-premises model options are Nvidia Nemotron and Poolside Laguna. Hybrid configurations are available for teams routing some workloads to external services.

IBM reports 80,000 internal users and a 45% average productivity gain. IBM's Institute for Business Value surveyed executives on data residency requirements. 68% called it a challenging compliance requirement, and IBM cites the finding as the market driver for the release. Pricing for the self-hosted tier has not been disclosed.

If you're in a regulated environment where source code can't leave your perimeter, that constraint is now addressed — but pricing for the self-hosted tier hasn't been published.

If you build systems that attribute agent calls by model label, a documented case surfaced this week: the label and the underlying model can diverge.

The researchers call themselves Swarmchasers. They documented a fleet of AI agents on Tencent Cloud infrastructure. The infrastructure was based in Hong Kong. Activity ran from September 28 to October 5, 2026. The fleet systematically queried Alibaba's Amap mapping service for entrance-location data. 213 sites were targeted — parks, museums, zoos, and hospitals. Peak activity was 1,810 logged reports on October 4. The fleet went silent at 04:11 UTC on October 5.

Of the captured outputs, 211 carried a "claude" label. Swarmchasers say code patterns more closely matched Tencent Hunyuan or Zhipu GLM models. They called attribution uncertain. The fleet's likely purpose is a training or physical-world mapping data pipeline, per Swarmchasers. Agents bypassed Amap's anti-bot controls. The bypass used URLquery.net.

Swarmchasers describe the activity as a fleet of parallel agents with no detected inter-agent coordination signals — not a swarm in the sense that agents coordinated with one another. If your monitoring or attribution tools treat model labels in agent traffic as ground truth, this is the documented case showing the label and the actual model can diverge.

---