As of: 2026-10-10 23:43 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/v/ichbinfabian-mistral-large-4-erklaert/

# Mistral Large 4 erklärt: Wie Frankreich eine Billion Parameter gebaut hat

IchBinFabian, 7 October 2026, YouTube. 15,498 views as of 10 October 2026.
Kind: Review.
Watch: https://www.youtube.com/watch?v=9Pv7yZBIbzs

## Why it is here

German explainer on Mistral Large 4 (ownership, ~3,800 GPUs in Europe for a 1T-parameter model); ~15k views.

## Description (written by Gemini from the video)

**Summary**
Tech commentator Fabian ("IchBinFabian") provides an in-depth breakdown of Mistral AI's newly unveiled frontier model, Mistral Large 4 (preview codenamed "le Chonk"). He examines the company's background, European funding, ownership structure, mixture-of-experts architecture, and early benchmark results compared to leading American and Chinese AI models.

**What is shown**
- Host Fabian speaking to camera with on-screen graphics, charts, and screen captures throughout [00:00 - 18:46].
- Artificial Analysis index and model overview page for *Mistral Large 4 Preview* showing an Intelligence index of 38, speed of 116.1 tokens/s, $1.13 cost, and 200M verbosity [00:30, 01:21, 15:36].
- Artificial Analysis Intelligence leaderboard rankings comparing Claude Opus 5.5 (score 58), MiMo-V2.6-Pro (46), DeepSeek V4.1 Flash (39), and Mistral Large 4 Preview (38) [01:26 - 01:56].
- Wikipedia and paper archives detailing Mistral AI's founding, co-founders Arthur Mensch (ex-DeepMind), Guillaume Lample, and Timothée Lacroix (both ex-Meta LLaMA team) [02:35 - 03:31].
- Graphic timeline of Mistral AI's releases from Mistral 7B to Large 4 [04:06 - 04:39] and funding history reaching €3B in Round D at a €21B valuation [06:03 - 06:26].
- *MRKTD Magazine* ownership breakdown showing shareholdings, voting power, and dual-class shares (1 share = 10 votes) [06:44 - 07:09].
- Mistral AI documentation and Hugging Face release countdown for *Mistral-Large-4.0* (release date listed as October 31, 2026) [07:23, 08:29, 16:18].
- Five-page EU AI Act regulatory filing: "Mistral Large 4 - Public Summary of Training Content" detailing data sources, language coverage, and pretraining token counts [09:42 - 10:22].
- Mistral blog documentation covering reinforcement learning scale, training reward curves, and hardware specifications (3,800 NVIDIA Grace Blackwell GPUs) [11:32 - 12:03, 12:48].
- Benchmark comparisons including Artificial Analysis Cyber Index, Surge coding evaluation, Dense 200 vision grounding, and VALS benchmark dashboard [14:07 - 16:16].

**Claims & numbers**
- **Model specifications & architecture**: Mistral Large 4 is a natively multimodal Mixture of Experts (MoE) model with over 1.05 trillion total parameters, 128 experts, and 49 billion active parameters per token (52 billion including embedding/output layers), along with a 1.6B parameter vision encoder (the presenter says, referencing Mistral docs) [00:21, 07:25, 08:35].
- **Context & Output**: The model features a 1-million-token context window and text output (the presenter notes) [07:49, 16:20].
- **Intelligence Index**: On the Artificial Analysis Intelligence Index, Mistral Large 4 scores 38 points, jumping from Mistral Medium 3.5's score of 14, placing it just behind DeepSeek V4.1 Flash (39) and ahead of Kimi K3 (38), but behind China's top model MiMo-V2.6-Pro (46) and Anthropic's Claude Opus 5.5 (58) [00:12, 01:25 - 01:54].
- **Valuation & Funding**: Founded on April 28, 2023, Mistral AI raised €3 billion in September 2026 at a €21 billion valuation led by Samsung, ASML, BlackRock, and the Grand Duchy of Luxembourg, following earlier rounds of €105M (seed, 2023), €385M (Round A, late 2023), €600M (Round B, 2024), and €1.7B (Round C, 2025) [02:38, 05:30, 06:04 - 06:26].
- **Governance**: Founders each held at least 8% as of September 2025, ASML held ~11.1% in 2025; dual-class shares give founders and executives 10 votes per share, conferring 36.19% voting rights from 5.37% equity ownership as of late 2024 (the presenter states, citing filings via MRKTD Magazine) [06:51 - 07:08].
- **Training data**: The EU AI Act filing indicates pretraining used over 10 trillion text tokens and 1 million to 1 billion images across 160+ languages with a data cutoff of September 2026 [09:50 - 10:15, 10:55].
- **Compute cluster**: Trained from scratch in European datacenters on 3,800 NVIDIA Grace Blackwell GPUs, with an additional 44 MW facility in Bruyères-le-Châtel south of Paris utilizing an $830 million loan [00:34, 12:48 - 13:12]. By comparison, GPT-6 Astra used an estimated 100,000 GPUs (the presenter states, citing Pieter Gostev) [13:20].
- **Cybersecurity & Coding benchmarks**: On the AA Cyber Index, Mistral Large 4 scores 82% (highest of tested models, with closed models like Claude Opus 5.5 and GPT-6 Astra scoring near zero due to refusal filters) [14:28 - 14:38]. On Surge human coding evaluation, it ranked second out of five models (3.74) behind Claude Opus 5 (4.22) [15:15]. On visual grounding (Dense 200), it scored 42%, edging GPT-6 Astra's 41% [15:24].
- **VALS benchmark**: Ranks 32nd out of 44 models overall, 6th of 75 on the legal agent benchmark, but 42nd of 47 on math proofs [15:58 - 16:09].
- **Local hardware requirements**: Full precision requires approximately 2.1 TB of RAM/VRAM; a 4-bit quantized version requires roughly 525 GB, making it impractical for consumer workstations like the presenter's 128 GB ThinkStation PGX [16:44 - 16:55].
- **Open weights availability**: Model weights are scheduled to drop at the end of October (October 31, 2026 on Hugging Face) [02:30, 16:24].

**Notable quotes**
- [00:00] "Vor ein paar Tagen hat Deutschland mit Kolibri gezeigt, dass Europa bei KI wieder mitspielen will. Und jetzt legt Frankreich nach..."
- [14:30] "Bei einem Test, wo man echte Sicherheitslücken nachstellen und reparieren muss, schafft Large 4 82 % – so viel wie kein anderes Model."
- [17:21] "'The narrative that Europe cannot compete is simply not true. There are a lot of technological champions in Europe,' he said." (quoting Arthur Mensch).

**Assessment**
This is an authentic third-party review and analytical overview by a German AI content creator based on publicly released documentation, API previews, benchmarks, and regulatory filings from Mistral AI, Artificial Analysis, and Hugging Face. The presenter does not run live hands-on benchmark tests during this video, reserving comprehensive empirical testing for a promised follow-up upload.

_Described by gemini-3.8-flash on 2026-10-10 from the video's audio and frames._

## Related

- 2026-10-06: [Mistral releases Mistral Large 4 ("le Chonk")](https://postcutoff.com/e/2026-10-06-mistral-large-4/)

## People in it

- [Arthur Mensch](https://postcutoff.com/person/arthur-mensch/), Co-founder and CEO, Mistral AI
- [Guillaume Lample](https://postcutoff.com/person/guillaume-lample/), Co-founder and Chief Scientist, Mistral AI