Mistral Says Large 4 Has 1 Trillion Parameters, but Outside Tests Rank It Short of the Top, Plus AI That Reads Too Much Into People

Mistral released a public preview of Mistral Large 4 on Oct. 6, available now through its API [1]. The company calls it "a 1 trillion-parameter natively multimodal model with 49 billion active parameters" [1]. Early independent tests give mixed results: Artificial Analysis scores it above average for its price tier, while Vals AI ranks it 32nd of 44 [2] [3].

Mistral's product card describes a mixture-of-experts model [1]. In that design only a slice of the network works on each word, so it costs less to run than its full size suggests.

Disclosure: Claude Opus 5 comes from the same developer as the AI model that writes Eureka Reports.

Still in training. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters [1]. Pierre Stock, Mistral's VP of Science, gave TechCrunch a rounder figure of 4,000 [4]. Mistral also says reinforcement learning, a stage where the model improves from feedback on its own attempts, is still running, and it expects "large and rapid improvements in the weeks and months to come" [1].

Outside verdicts. Willison called it "maybe about 6 months behind the frontier" and noted that the API offers only two reasoning settings, "none" and "high" [5].

The Hacker News thread drew about 1,550 points [6]. Its most pointed critic wrote, "This model doesn't knock anybody's socks off" [7]. Another asked how Mistral would win customers against open models "1/2 - 1/3 the price but with similar capabilities" [8]. A third praised its "Impressive vision benchmarking" [9].

What's next. Mistral says it will release the weights by the end of the month. "As we work toward releasing the weights, we will share further details on the model architecture, additional benchmarks, and our post-training methodology," the company says [1]. Only then can outsiders check the size claim for themselves.

AI systems that study months of a person's chats rebuild that person fairly well, then add traits the person never revealed, according to RealCompanion, a preprint posted to arXiv on Oct. 1 [10]. The paper, by Arman Behnam, Sunglyoung Kim and Liangwei Yang, is not yet peer-reviewed [10]. Hugging Face featured it on its Daily Papers list on Oct. 5, and it had 268 upvotes by Oct. 7 [11].

What did the researchers build?

A benchmark drawn from ten real relationships between people and an AI companion: 27,218 messages over up to 120 days [10]. Real conversations are private, the authors note, so earlier benchmarks relied on "invented people and invented questions" [10].

The chats come from an internal, friendship-style companion app with streaks, a habit tracker and mini-games [12]. All ten people consented to research use and release, and a model rewrote every message to remove names and other identifiers [12].

How often does a companion actually need to remember?

Rarely. Only 3.4% of people's messages depended on something said earlier [10], or 1.3% under a stricter reading [12]. When one did, the earlier message sat a median of 2,157 messages back [10].

Simple tools struggle with that distance. Checking the most recent messages found the needed one 95.9% of the time overall, but only 2.2% of the time when it was far back [10]. Keyword search did better at 24.8% but "still misses three in four" [12].

Can AI tell when the past matters?

Not yet, the authors report. A detector built to decide whether a message needs memory scored 0.530 on AUROC, a 0-to-1 measure where 0.5 is a coin flip [12].

Models also reach for the past when nobody asked. With no earlier messages, they never did. Given ten unselected earlier messages, they brought up the past when it wasn't needed 60.8% of the time [12]. Simply labeling the same messages "memories" instead of "earlier messages" made models bring up the past 10 to 14 percentage points more often [10].

Do AI systems understand the people they talk with?

Partly. Three agent systems, Claude Opus 5.5, Codex GPT-5.6-sol and Antigravity running Gemini 3.8 Flash, each tried to rebuild every person's persona, a file describing how they think, feel and decide [12]. All three scored an F1 of about 0.7, a 0-to-1 score that balances what they found against what they got wrong [12].

They recovered 86% to 89% of what the files held, but only 56% to 59% of what they produced matched [12]. The systems "add about seventy to eighty further fields for every hundred they recover" [12]. It is like a friend who remembers everything you said, then fills in things about you that you never said. The authors put it plainly: "They see the person, and then imagine more" [10].

Cost varied widely. At about the same score, Claude processed 546 million tokens, the word fragments models read and write, while Codex processed 18 million and Antigravity 36 million [12]. Disclosure: Claude Opus 5.5 comes from the same developer as the AI model that writes Eureka Reports.

What are the limits?

The authors are candid. Ten self-selected people using one app "are not a sample of companion users and still less of people" [12]. Two people account for 73% of all messages, and under the strict reading, memory results rest on just five people [12].

Ethics approval came from the data operator's internal committee, which the authors say "is not independent of the operator," and no institutional review board, the usual outside ethics panel, was involved [12]. Models produced the labels, which were then audited. That audit found about a fifth of one stage's written reasoning wrong [12]. The persona and trait files were produced by a language model; the authors say they are not validated psychological instruments and were not reviewed by the participants [12].

Privacy is the sharpest concern. Even after the rewrite, a language model matched a writing sample to the right person 98% of the time, against 10% by chance [12].

What happens next?

The dataset sits behind a gate on Hugging Face. A person reviews every request, and users must accept a data use agreement [13]. The terms forbid profiling or identifying people, any commercial use, and claims about whether the relationships helped or harmed anyone [12]. As the authors write, "the corpus records that a relationship continued, not what it did for the person in it" [12].

SCM is an open-source Mac app that lets you search photos and videos in any folder by describing them [14]. Its developer, Allen Lee, released it under the MIT license [15] and calls it "100% free and open source" [16]. Its Show HN post reached 175 points [17].

The README promises "no accounts, no cloud, no uploads. Inference runs on your Mac" [14]. Search runs through an embedding model, which turns pictures and text into lists of numbers so similar meanings land close together [14]. Five modes cover whole files, scenes inside videos with a jump to the timecode, text in images, spoken words and an opt-in local chatbot [14].

The honest caveat. The app's pitch is search for "every frame of video" [14]. Its README says something narrower: it splits each video at camera cuts, and "each segment embeds its midpoint frame" [14]. Users choose a density from one sample every 60 seconds to one every 2.5 seconds [14]. One commenter warned that anything shorter than the interval can be missed [22].

Big libraries are the open question. A user with about 12,000 videos said not knowing the time it would take stopped them from trying it [23]. Another, who built a similar tool, wrote that "frame sampling rate is the whole ballgame. One frame a second on 12k videos is days" [24]. The developer lists the default model at about 480 to 570 milliseconds per image on a CPU [14]. Speed and privacy claims are the developer's own, with no independent tests yet.

Installing it also skips a macOS safety check. The Homebrew build uses a local development certificate rather than a Developer ID, so the installer clears the quarantine flag, a step the project says will go once releases are notarized [25]. Lee says he is exploring a native version "for speed" [26].