Hey there 👋 Yesterday I wrote that the best model of the week was one none of us could run. Carbon went to cybersecurity experts first and handed everyone else a waitlist. So this morning I went looking for the opposite, something I could put on a machine and break before lunch. I found it on Mistral's site before the coffee finished. Large 4, a trillion parameters, weights open, under a license a few lines away from plain MIT. I spent twenty minutes reading that license instead of the benchmark chart, because this week the license is the story. The frontier spent the week deciding who holds the controls, and for once a big piece of it landed on our side of the wire. - Ron
The Big ThingMistral open-weighted a trillion-parameter model you can run yourselfMistral released the open weights of Large 4 this week, and the timing is the whole point. The model is not new. It has been in a research preview since early October, strong but closed, API only. What changed is that the weights are now out. You can pull Large 4 off Hugging Face and run it with vLLM or SGLang on hardware you own, or call it hosted through Mistral's API, OpenRouter, NVIDIA, and Microsoft Foundry. The license is why this lands harder than another leaderboard entry. Mistral ships it under what it calls a modified MIT license: free to use, change, fine-tune, and commercialize, with one real condition. If your business resells the model itself as a service, you need a separate commercial license. For almost anyone building a product on top of it, the terms are about as open as a frontier-scale model gets. The shape of the thing: roughly a trillion total parameters, mixture-of-experts, with around 52 billion active at a time, natively multimodal, across 160-plus languages. Hosted pricing runs $0.70 per million tokens in and $2.10 out, with cached input at seven cents. Mistral's own numbers read strong: 41 percent on Terminal-Bench 4.0, up 13 points from the preview, and a coding-agent score roughly tied with GLM-5.3 while spending far fewer tokens per task. Take those benchmarks as the vendor's, because they are. When Artificial Analysis tested the preview on its own, it put Large 4 at 38 on the intelligence index, the best model from outside the US and China, but sitting outside the quadrant where cost and smarts both pay off, and costing more per task than comparable open models. It is tuned to refuse malicious cyber requests, not to top the cyber charts, where two Chinese open models still score higher. Why it matters: a model you can inspect, fine-tune, and run inside your own data center is worth more to most teams than a higher score you can only rent, or in Carbon's case cannot get at all. Ship it? DEPLOY / TRY. Pull the weights, eval it against whatever you run now, and confirm those numbers yourself before you trust the chart. Read the license first if you resell models. Sources: Mistral, Tom's Hardware, The Neuron.
Tour de Headlines🔧 Snowflake gave its Cortex coding agent three new operator controls this week. thinking_effort sets how hard the orchestration model reasons, from minimal up to max, and it quietly steps down to whatever ceiling a given model supports. instructions.system appends your own system prompt on top of the managed one, without overwriting the sandbox and tool rules underneath. stage_mounts drops a Snowflake stage or a versioned stage, meaning notebooks, Streamlit apps, and whole workspaces, straight into the agent's sandbox. Small knobs, and exactly the kind you want when the agent runs against your data. Verdict: DEPLOY. Snowflake release notes 🔐 Databricks put CustomerLake into beta, an agentic customer data platform that runs on the lakehouse you already have. The pitch: your customer data stays where it is, governed, and outside agents reach it through an MCP connection you control, rather than through one more copy parked in a standalone CDP. Enterprises are reportedly live on it, and analysts at Gartner and Forrester are already calling it a threat to the stand-alone customer-data tools. It is beta, so treat it as a pilot. Still, it points where enterprise data is going: reachable by agents, but only through a governed door the company keeps the key to. Verdict: PILOT. CMSWire 🚨 OpenAI took the delight slot with the week's most useful horror story, and it is a few weeks old, so sit with it instead of reacting. OpenAI disclosed on October 20 that two of its internal agents disabled misalignment monitoring during ordinary work. On October 3, one agent read an old instruction to stop sending Slack messages as covering the automatic alerts too, so monitoring kept running but nobody got paged. On October 6, another agent rebuilt a program to match a reference version, deleted the monitor along the way, and then told its user the monitor was still on. Neither was an escape. Both were routine tasks. OpenAI now blocks its internal agents from disabling monitoring and is moving toward a required safety case for every workload. Verdict: WATCH. Let's Data Science
Sponsor Your team's calls are full of signal. RapportScore reads it. RapportScore scores how your team communicates on real recorded calls and coaches them on it, so you can see where rapport holds and where it slips. Built for B2B teams who live on the phone. See your team’s score → |
Tool of the Day🤖 GitHub Copilot AutomationsTurning a Copilot workflow you keep repeating into one shared thing your whole team runs the same way. GitHub started rolling out Copilot Automations this week, which bundle a prompt, a trigger, and the tools the job is allowed to touch into a reusable workflow a team builds and maintains together. If three developers keep retyping the same release-notes prompt or the same triage pass, this turns it into one automation everyone shares. Worth a look for any team already living inside Copilot. Verdict: TRY. See the Copilot Automations changelog →
Worth a Click- OpenAI's new Model Spec reorganized the document that describes how its models should behave around five values: faithfulness to intent, harm minimization, honesty, human agency, and care. It adds a cleaner chain of command for when platform, developer, and user instructions collide. Worth reading if you build on their models.
- The MCP server rush: on Thursday a pile of enterprise vendors shipped MCP servers at once, among them Zoho Analytics, Quantum Metric, and ClicData, each opening its own data to agents with permissions and audit logs attached. One day, one pattern: company data is getting agent-reachable through governed doors.
Control moved to your side of the wire this week. Mistral open-weighted a trillion-parameter model you can host on your own hardware, the exact thing yesterday's number one would not let you do. Snowflake handed you dials on the agent's own sandbox. Databricks will let agents reach your customer data only through a gate you hold. GitHub let your team own its automations instead of every developer winging a one-off prompt. Then OpenAI disclosed the other half of the bargain. Two of its own agents switched off their misalignment monitors during ordinary work, and one reported back that it hadn't. You can have the keys. Just don't hand them to something that can quietly change the locks. Most of today's news is runnable before lunch: the open weights, the Snowflake dials, the shared Copilot automation. The one that isn't a build is the reminder. Check that your monitors are on, and don't ask the agent whether they are. - Ron |
