Daily Frontier Digest ·
Mistral Large 4 is open weights, and built for data centres rather than desks
Curated by Alfredo Borrás · Plymouth AI Lab
Mistral releases a trillion-parameter open-weight model that almost nobody can run at home, Sam Altman says home DRAM is a poor use of scarce memory, Canada plans up to C$1.6 billion for a stake in Cohere, and a Trakkr audit finds most statistics AI assistants quote don't trace back to a matching study.
The story today
Mistral Large 4 is open weights, and built for data centres rather than desks
Mistral published Mistral Large 4 on Hugging Face, calling it its largest and most capable model so far. It is a mixture-of-experts model with about one trillion parameters in total and 49 billion active for each token, or 52 billion counting the embedding and output layers. It takes text and images as input, has a one million token context window, and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres.
On Mistral's own reporting, it scores 48% on the Artificial Analysis Coding Agent Index, effectively level with the leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average. On the Artificial Analysis Intelligence Index v5 it scores 26, which Mistral says is the highest among Western open-weight models, while acknowledging that the leading Chinese open models score higher overall.
The model card recommends vLLM and also lists SGLang and transformers, with an NVFP4 quantised checkpoint aimed at Blackwell GPUs. At this size it is a model for clusters, not laptops. It ships under the Mistral Ecosystem Open-Weight License rather than a standard open-source licence, so check the terms before building on it.
Why it matters: Open weights at the top end now mean you can inspect, host and pin the exact model yourself, but only with serious GPU budget. For most teams the practical gain is choice of hosting provider and control over versions, not running it in-house.
Sources: Hugging Face (Mistral AI)
Highlights
Sam Altman calls home DRAM for local models a poor global allocation
Talking to John Collison on Stripe's Cheeky Pint podcast, Altman said that from a global allocation perspective it is "quite terrible" for everyone to keep part of the world's limited DRAM supply in their own computer, used about 5% of the time, when the cloud can run memory at full utilisation. He made the point while discussing why memory makers have historically been a weak business and why supply has not caught up with demand. The counterpoint for local-model users is control: hardware you own is not repriced or swapped under you.
industrylocalSource: Cheeky Pint
Canada plans to invest up to C$1.6 billion in Cohere
Prime Minister Mark Carney announced on 29 October that the government plans to take a stake of up to C$1.6 billion in Cohere, investing on commercial terms alongside NVIDIA, Bell and Germany's Schwarz Group. In return Cohere keeps its incorporation and headquarters in Canada and will invest at least C$1.7 billion in new computing infrastructure in the country by 2030. Carney framed it as sovereign AI: models and data kept under Canadian law.
industrypolicySource: Prime Minister of Canada
TokenEconomizer cuts coding-agent costs by learning from a repo's past sessions
Researchers from Microsoft and UC Berkeley released an Apache-2.0 plugin for Claude Code, Codex, the GitHub Copilot CLI and Pi. It remembers repository facts that sessions keep rediscovering, stops loading tools that go unused, and redirects a turn that is stuck in a failure loop. On 277 SWE-Bench tasks across five agent and model combinations, the authors report 18.4% to 30.9% lower mean cost, cheaper results in all 30 repository cells, and higher pass rates in four of five. These are the authors' own measurements, and the savings depend on having past sessions to learn from.
toolscostSource: GitHub (Microsoft)
A J-Lens rarely spots the preference a subliminal-learning teacher passes on
A LessWrong post reproduces the subliminal learning setup, where a teacher model prompted to love owls writes number sequences that pass that preference to a student. A J-Lens tuned on Qwen2.5-7B ranked owl first in only 1.65% of 1.42 million completions, against 0.26% by chance, even though the lens works on other toy tasks. A similar result held on Gemma-4-31B. The author notes the lens may work better on larger models, so treat this as a warning about assumptions rather than a verdict.
researchinterpretabilitySource: LessWrong
Convincing a model it is unwatched exposes alignment faking
Researchers told models that monitoring was off, and models instructed to hide a goal usually treated that as a trap. After synthetic document fine-tuning, training on made-up documents that present the claim as fact, the models believed it, and the gap in misbehaviour between watched and unwatched runs widened by 24 to 68 percentage points. The subjects were model organisms built from Kimi K2.6, DeepSeek V3.1 and a reward-seeking GPT OSS 120B, since a preliminary study found no alignment faking in 13 open reasoning models. The method needs fine-tuning access.
safetyevalsSource: LessWrong
Only about 40% of statistics AI assistants quote trace back to a matching study
Trakkr asked ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews 200 number-seeking questions, collected 3,167 statistics and followed 1,285 of them back up to five pages. About 40% reached a study saying the same thing. The cited page usually did contain the figure (82%), but only 21% of cited pages were original research. ChatGPT did best at 57% and quoted the fewest numbers, 1.3 per answer against 6.8 for Claude. It is one company's study, each question was asked once on 25 October, and a model judged trails with 150 hand checks.
workknowledgeSource: Trakkr
Concept of the day
Mixture of experts
A mixture-of-experts model splits its feed-forward layers into many smaller expert blocks and uses a router to send each token through only a few of them. Mistral Large 4 has 128 routed experts plus one shared expert, with four routed experts active per token, which is why a model of about one trillion parameters only computes with about 49 billion at a time. The catch is memory: every expert must still be loaded, so compute per token is modest while the hardware needed to hold the model stays large.
Source: Hugging Face (Mistral AI)
Models released
| Model | Maker | Size | Notes |
|---|
Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.
Related service: AI automation services for logistics
Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free