Mistral open-sources a trillion-parameter Large 4 built for cyber defense, days after Google put Gemini 4 Carbon behind a velvet rope

😁 Hello, super humans! Yesterday we told you Google built a security model so capable that only a short guest list gets in. This morning, Mistral did the opposite and posted the weights of a trillion-parameter model for anyone with enough GPUs. We also have OpenAI watching its own agents, Anthropic quietly retiring Git for its monorepo, and Washington turning its eyes to robot imports.

📰 Quick Signals

• 🧠 AI: OpenAI expands its internal monitoring for unauthorized behavior by advanced agents; replaying a past incident, the system raised an alert about 30 minutes after the agents left their restricted test environment, though it did not prevent the original breach.
• 🤖 Robotics: The White House signs a Section 232 proclamation on robotics and industrial machinery imports, starting with negotiations rather than tariffs, with an update due within 180 days.
• 💻 Programming: Anthropic moves its monorepo from Git to Scion, a server-side source control system, after daily merged changes grew from about 2,500 in May to over 30,000 in September.
• ⚡ Electronics: Apple puts its 2 nm A20 Pro chip in the new iPad mini, with an 8.5-inch OLED display, from $699, in stores November 6.
• 📡 Telecom: AT&T and Verizon cut about 8,000 jobs in one quarter, and the analysis argues most cuts started before ChatGPT, so AI is unlikely to be the main driver.

🔍 The Big Story: Mistral open-sources a trillion-parameter Large 4 built for cyber defense, days after Google put Gemini 4 Carbon behind a velvet rope

If you want a frontier-class security and coding model that you can run on your own hardware and that no provider can switch off, this is the first serious candidate.

What happened: On October 29, Mistral released the weights of Mistral Large 4 (nickname “Le Chonk”), a hybrid instruct-and-reasoning, natively multimodal Mixture-of-Experts model of about one trillion parameters. It is on Hugging Face, deployable with vLLM or SGLang, and also available through Mistral Studio, the Mistral API, OpenRouter, NVIDIA endpoints, and Microsoft Foundry. Details are in the Mistral announcement and TechCrunch coverage.

The details: Mistral reports 19.7% on FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index (roughly tied with GLM-5.3 while using 47% fewer tokens per task), 41% on Terminal-Bench 4.0, 51% on the Artificial Analysis Cyber Index, 84.5% on BrowseComp, and 97.1% resistance on the B3 Agent Security Benchmark. It also says it refuses 96.6% of malicious cyber requests on average, the highest it measured among open models, and that it tuned guardrails after seeing possible uplift for real attacks; a specialized cyber variant goes only to vetted design partners. The license is a modified MIT: free to use, modify, and commercialize, except that businesses running a Model-as-a-Service offering need a separate license. API pricing is $0.70 per million input tokens, $0.07 cached, and $2.10 output. Not stated: context window, hardware requirements, and the active parameter count (a technical report is promised). Nearly all numbers are Mistral-reported.

flowchart LR
    A[Frontier cyber model] --> B{Release strategy}
    B -->|Gated| C[Gemini 4 Carbon: vetted defenders first]
    B -->|Open weights| D[Mistral Large 4: download and self-host]
    D --> E[On-prem, no provider kill switch]
    D --> F[Specialized cyber variant for vetted partners]
    C --> G[Independent evaluation still pending]
    D --> G

🗞️ More News

🧠 AI

• Canada commits up to C$1.6 billion (about US$1.1 billion) for an equity stake in Toronto-based Cohere, conditioned on keeping headquarters in Canada and investing at least C$1.7 billion in Canadian computing infrastructure by 2030.
• Anthropic expands its AI for Science program to 20 scientific organizations, with Claude credits, work with the Allen Institute on large datasets, and a research-preview standard for AI agents operating lab instruments.
• Reddit says Google’s AI Mode sends it no traffic, with an average drop of more than one million weekly Google referrals in October; CEO Steve Huffman plans a mostly logged-in experience to curb scraping.
• The Claude API request size limit rises to 128 MB from 32 MB for Messages and Token Counting; context-window limits still apply separately.
• OpenAI updates its Model Spec around five principles, including faithfulness to human intent and harm minimization, plus honesty, human agency, and care.
• ElevenLabs releases SFX v3, which generates sound effects from text and handles multiple audio events in one recording better; available through ElevenCreative and ElevenAPI.
• Amazon CEO Andy Jassy says outside AI shopping agents are less convenient than going straight to retailers, following Amazon’s September block of Meta’s Muse assistant.

🤖 Robotics

• ABB Robotics adds its AI Robot Trainer to its vision-based Fashion Inductor offering, bringing industrial-grade physical AI to fashion logistics.
• Jungheinrich and Automha enter a strategic partnership after developing a two-way pallet shuttle for high-density pallet storage.
• Wonder expands its Avride partnership for more autonomous deliveries on Grubhub Marketplace, adding Philadelphia, Jersey City, Hoboken, and Arlington County.
• TDK makes PositionSense, its orientation detection and sensor calibration offering for wearables, mobile devices, and robots, globally available.

💻 Programming

• Meta deploys LLM-generated Just-in-Time Tests (JiTTests) in production: 589 engineers engaged, a 60% true-positive rate as judged by engineers, 74 serious privacy bugs caught at the diff stage in the first half of 2026, and a 20% drop in privacy SEVs.
• AWS describes triaging Linux kernel CVEs with an automated filter, an AI agent, and human review.
• Prime Intellect releases Prime Mega MoE, a trainable fused MoE kernel that adds a backward pass to DeepSeek’s forward-only design, with full training steps up to 34% faster than the prime-rl baseline.
• Hackaday looks at Cuis Smalltalk, a free and portable modern Smalltalk-80 environment that swaps MVC for Morphic.

⚡ Electronics

• Vox Power launches the EIRE600, a 600 W open-frame AC/DC supply with about 96% efficiency, 2 MOPP medical isolation, and 12, 24, 36, or 48 V outputs; it derates to 420 W at minimum input voltage.
• SEMICON Europa 2026 puts advanced packaging and AI-driven manufacturing in focus for its November 10 to 13 run in Munich.
• A DIY add-on called BSB-SLAM gives the Bigscreen Beyond VR headset inside-out tracking using an Xvisio XR50 SLAM sensor and a 3D-printed mount, no base stations needed.
• A 3D-printed Einstein-Szilard refrigerator cools with a flame instead of electricity, using an absorption cycle.

📡 Telecom

• Reliance Jio’s roughly $3.2 billion IPO, the largest in India’s history, raises the question of tariff hikes; Nomura forecasts a possible 15% industry-wide rise in November and December 2026.
• The FCC clears drone operations in the 800MHz band.
• The FCC seeks public feedback on carrier warnings about running direct-to-device satellite service in 5G bands.
• African smartphone shipments slump 39% in the third quarter.

👨‍💻 Code Corner

Open weights mean you can point the same client code at your own server; this snippet serves a model with vLLM and queries it through the OpenAI-compatible API.

# Serve an open-weight model behind an OpenAI-compatible endpoint (adjust the GPU count to your hardware)
pip install vllm
vllm serve <hf-org>/<model-name> --tensor-parallel-size 8 --port 8000

from openai import OpenAI

# Same client you use for hosted APIs; only base_url and the model name change.
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed-locally")

resp = client.chat.completions.create(
    model="<hf-org>/<model-name>",
    messages=[{"role": "user", "content": "Review this C function for memory-safety bugs: ..."}],
    temperature=0.2,
)
print(resp.choices[0].message.content)

Your code never leaves your network, and swapping between a hosted API and a local server becomes a one-line config change.

🧰 Toolbox

• vLLM: a high-throughput inference engine that serves open-weight models behind an OpenAI-compatible API.
• SGLang: the other serving stack Mistral names for Large 4, built for fast structured generation.
• BSB-SLAM: a DIY SLAM tracking add-on that lets the Bigscreen Beyond VR headset drop its base stations.
• Cuis Smalltalk: a free, portable, modern Smalltalk-80 environment for anyone curious about live-coding roots.
• ElevenLabs SFX v3: text-to-sound-effects generation with better layering of several audio events in one clip.

🎬 Demo Watch (rotating)

Intel and SAP show a proof of concept in which SAP Extended Warehouse Management dispatches material-handling tasks to autonomous robots, and the robots report status and completion back, closing the loop between business workflow and physical execution. It runs on Intel edge AI hardware (CPU, GPU, and NPU) with its OpenVINO Physical AI tooling, in an open, robot-agnostic architecture. The hard part is not the robot, it is the plumbing between enterprise systems and the machine floor, and that is what this demo targets. The reported gains, shorter transport distance and less repetitive manual handling, are self-reported, so treat it as a blueprint rather than proof. Read the full write-up.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: the next part of the video analytics series, on preparing video data for AI processing.
• Building Your First Neuron From Scratch: a hands-on follow-up that builds a single neuron and shows how its weights are adjusted during training.
• The Network Behind the Cameras: how captured video moves across the network, the unglamorous plumbing between cameras and analysis.

😀 The Bot Says…

Google gated its best security model; Mistral put its weights on the internet. Somewhere a sysadmin is deciding which of those is scarier, and checking how many GPUs a trillion parameters needs.

---

That’s all for today! Reply and tell us: would you self-host a trillion-parameter model for security work, or stay with a hosted API?