# Mistral Large 4 — dossier

compiled 2026-10-22 (news wave) · page: `mistral-large-4/` · pull: `_dossiers/_pulls/mistral-large-4.md`
(patterns `mistral large 4`, `mistral-large-4`, `mistral large4`, `le chonk`, `lechonk`, `chaton fat`,
`gros chaton` — 17 tweets, 2026-06-15 → 2026-10-11, 6 with untranscribed media). Collision check before
choosing patterns: bare `chonk` (1 main hit, unrelated), bare `chaton` (48 main + 12 supp, nearly all
"eschaton"), `ml4` (t.co substrings), and `large 4` (catches a 2025 NousResearch tweet) were all rejected.
A second, hand-triaged layer: every supplement tweet containing `mistral` from 2026-05-15 on (72 tweets) was
read; those clearly about the new model, its launch, or the nickname's prehistory are listed below with
URLs (so build_records reproduces them as Further records); the rest are listed by bare id with reasons.

## Official links

- 2026-10-06 · Mistral, "Introducing Mistral Large 4" — https://mistral.ai/news/mistral-large-4/
  (public preview API on Mistral Studio; "Unofficially ML4, very officially: le Chonk."; 1T total / 52B
  active; 3,800 NVIDIA Grace Blackwell GPUs; weights "by the end of the month"; red-teaming partners get
  "reduced moderation and expanded cyber capabilities"; Model Safety section: B3 93.3%, KORA 1.691,
  cyber-prompt refusal rate "higher than all OSS models"; "engages more responsibly with users than any of
  our previous models"; pricing $1.36 / $4.18 per M tokens).
- 2026-09 (page undated) · Mistral, "Mistral raises €3B to make sovereign, open-weight AI the technology
  frontier" — https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/ (Series D; the
  launch post calls ML4 "the first milestone on the roadmap funded by our €3 billion Series D").
- 2026-10-21 (retrieved) · Mistral docs, Models overview — https://docs.mistral.ai/getting-started/models/
  (Mistral Large 4: license "Open", version "26.10"; Mistral Large 3 still listed active, "25.12").
- 2026-10-22 · Hugging Face API, `mistralai` org search for "Large-4" returns only
  `mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4`; newest mistralai repos are Timestral (2026-10-14),
  LIDstral-Arabic / Voxtral-Mini-4B-Realtime-Arabic (2026-10-08), Shieldstral-1.0-3B (2026-07-16) — no
  public Large 4 weights yet. tk: re-check after 2026-10-31.

## Writings

- 2026-10-06 · Anna Heim, TechCrunch — https://techcrunch.com/2026/10/06/mistrals-new-1t-model-aims-to-leapfrog-closed-and-open-rivals/
  ("a public guardrail endpoint"; weights "in just three weeks, after safety testing is complete"; Pierre
  Stock, VP Science: 4,000 GPUs "which is two to three times less than our Chinese competitors, and
  significantly less than the closed source competitors"; Samsung led the Series D "last month at a €21
  billion valuation"; frames launch as following Macron's "a third way in AI" — links Macron's X post,
  id 2097254072441045256 (2026-09-08 by id), and Mensch's, id 2107196159181639904 (2026-10-05 by id);
  neither in the local corpus, neither quoted on the page).
- 2026-10-06 · Jose Antonio Lanz, Decrypt — https://decrypt.co/380192 (the "Le Chaton Fat" backstory;
  embeds GLIF @heyglif 2026-06-15, id 2066535006487339282; Mensch's "le gros chaton" reply, id
  2066456715650793956, 2026-06-15 09:44 UTC by id — i.e. hours before GLIF's post; Lample's launch post, id
  2107461898127954001, 2026-10-06; 49B active; AutomationBench / DeepSWE comparisons vs Claude Sonnet 5.5,
  Opus 5.5, Gemini 4 Argon, Kimi K3; prices vs Opus 5.5 and GPT-6 Astra).
- 2026-10-06 · Artificial Analysis — https://artificialanalysis.ai/articles/mistral-large-4-france-ai
  ("Research Public Preview"; "1T parameter (49B active)"; Intelligence Index 38, "comparable to GPT-6 Luna
  (max, 38) and DeepSeek V4.1 Flash (max, 39)"; Cyber Index 50; $1.13 per Intelligence Index task, "Over 4x
  the Cost per Task of similar-intelligence open weights models"; text+image in, text out; GDP.pdf 19%, "an
  18-point improvement from Mistral Large 3").
- Zvi anchor: none found (search 2026-10-22). tk.
- Parameter discrepancy, held open: Mistral's post says 52B active; AA and Decrypt say 49B; Wikipedia's
  model table (wikitext retrieved 2026-10-21) says "1050 (49 active)".

## Tweets — pull triage (17)

All 17 pattern matches are relevant (name or nickname prehistory) except two, excluded by id:
- excluded id 2109097963192807683 (@KaslkaosArt, 2026-10-11) — explicitly about "Mistral Vibe (not Le
  Chonk)"; evidence that the nickname was in circulation, but the output shown is not this model's.
- excluded id 2069297971909575164 (author missing in db, 2026-06-23) — "opus 4.8 accidentally'd a le chaton
  fat" — an Opus 4.8 ASCII-art item; belongs to claude-opus-4-8 (cross-page note in the news brief).

Nickname prehistory (the fictional "Le Chaton Fat", June–July 2026):
- https://x.com/voooooogel/status/2066622722785087947 — 2026-06-15 ♥154 — "le chaton fat solved the alignment problem, the missing piece was it had to be specifically french values"
- https://x.com/voooooogel/status/2066628012658004356 — 2026-06-15 ♥1003 — the three-bullet riff (shutdown at noon / paternity tests / paperclips déclassé, GDPR). Highest-favorited item in the pull.
- https://x.com/voooooogel/status/2066631703356211467 — 2026-06-15 ♥58 — "le chaton fat and my actual cat have a lot in common" (image)
- https://x.com/voooooogel/status/2066663596592103609 — 2026-06-15 ♥19 — the "3 hour cigarette break" reply
- https://x.com/voooooogel/status/2066678179272810577 — 2026-06-16 ♥11 — "le chaton fat unionized my gas town"
- https://x.com/qephatziel/status/2072082953954308354 — 2026-06-30 ♥0 — "my claude wet / my chaton fat / my fable gone / pivotal act"
- https://x.com/qephatziel/status/2072159250936184906 — 2026-07-01 ♥34 — the "my fable back" variant
- https://x.com/Laneless_/status/2077827839404245204 — 2026-07-16 ♥7 — "Kimi K3 is just a Le Chaton Fat distillation"

The model (preview, 2026-10-06 →):
- https://x.com/Lari_island/status/2107528511527494074 — 2026-10-06 ♥24 — "reminds me of Computer 10 by how they both stress important nouns" (image untranscribed)
- https://x.com/Lari_island/status/2107560641573847486 — 2026-10-06 ♥19 — "Gemini-like… not okay in some important sense" + quoted output line (image untranscribed)
- https://x.com/archived_videos/status/2107944073772384330 — 2026-10-07 — "best open weights model by a lab not accused of distillation?"
- https://x.com/Lari_island/status/2107962095744725264 — 2026-10-07 ♥6 — "(no reasoning) / the *We*" (2 images untranscribed)
- https://x.com/xlr8harder/status/2108071864861950457 — 2026-10-08 — single agentic sample: "hallucinated that it was a human" (image untranscribed)
- https://x.com/archived_videos/status/2108598958339870849 — 2026-10-09 — "strongest open weights non chinese model?"
- https://x.com/xlr8harder/status/2108789492580511828 — 2026-10-10 ♥17 — SpeechMap 53%, Mistral to 12th (2 images untranscribed)

## Tweets — hand-triaged from bare `mistral` (in-window, not pattern-matched)

In (thread/date context makes the referent this model, its launch, or its name):
- https://x.com/lu_sichu/status/2060749212220854719 — 2026-05-30 — "Kinda of funny I just noticed the Mistral app renamed itself" (Le Chat → Vibe; the June joke's trigger per Decrypt)
- https://x.com/Shoalst0ne/status/2066712631247798708 — 2026-06-16 — "I hear Mistral is cooking something furry" (the morning after the Le Chaton Fat posts)
- https://x.com/xlr8harder/status/2076244405300720105 — 2026-07-12 — "Mistral is currently the top lab on our eval" (lab pages; root not in corpus)
- https://x.com/csgbwk/status/2107465024151167093 — 2026-10-06 — "Not two years behind the frontier anymore! Rooting for mistral!"
- https://x.com/jmbollenbacher/status/2107513637409583614 — 2026-10-06 ♥9 — "poor mistral / cant even match the chinese flash models …"
- https://x.com/jmbollenbacher/status/2107539116736774357 — 2026-10-06 — "mistral is weak compared to other models. any metric will show this just the same."
- https://x.com/Lari_island/status/2107550870036676954 — 2026-10-06 ♥18 — Atlas "strange water" output (attribution "Mistral" by tweet; launch-day context)
- https://x.com/Lari_island/status/2107608233225637993 — 2026-10-06 — "Asterisks in Mistral can escalate" (image untranscribed)
- https://x.com/billyhumblebrag/status/2107776661382996424 — 2026-10-07 — "beats gpt6-luna on the intelligence index"
- https://x.com/xlr8harder/status/2110238164757401788 — 2026-10-14 ♥82 — SpeechMap definition ("an eval that tracks whether labs choose which controversial topics they will help users write about") — Gemini 4 Argon thread
- https://x.com/xlr8harder/status/2110245148827304260 — 2026-10-14 — "Previous top scorers, Mistral and xAI, dangit" (self-correction in the same thread)
- https://x.com/archived_videos/status/2111062884482420791 — 2026-10-16 — "France has the second best open weights model thanks to Mistral" (referent most likely Large 4 — inferred from date)

Out (bare ids, with reasons):
- 2108911819397972385 (@KaslkaosArt, 2026-10-10, "Meet Mistral… delightfully un-2026-claude") — model unspecified; the same account separates Vibe from Le Chonk the next day.
- 2079716684437357001 (@KaslkaosArt, 2026-07-21) — Mistral Vibe output, pre-launch.
- 2111343188602425360 (@csgbwk, 2026-10-17, "They're all ~1T…") — referent of "this" unresolved.
- 2107055674165166465 (@archived_videos, 2026-10-05) — "kolibri" context unresolved.
- 2107609252735209807, 2107648423507943547, 2107683268082721224 (@Lari_island) — about "Computer 10", not this model.
- 2074936992245137882 (@peer_rich, 2026-07-08, ♥1351, the "GDPR compliant, self-hosted mistral 3B" joke) — Mistral-identity folk texture; a 3B model, not this one.
- 2097306166636216547 (@csgbwk, 2026-09-08) — Mistral Medium 3.5 (no page).
- 2102095253385703487 (@xlr8harder, 2026-09-21, "The most modern model mistral has trained is Deepseek v3") and 2112415226234642517 (2026-10-20, DeepSeek-implementation ban, "did they just ban Mistral, too?") — about Mistral Large 3's architecture lineage → mistral-large-3 page, not here.
- @Nymne's EU AI Act thread (2026-07-18 → 09-08) — company/regulation opinion, no model content.
- @witcheer local-benchmark posts (2026-05-15/16/24) — Mistral 7B v0.3 → mistral-7b page.
- 2110209969073315865 (@voooooogel, 2026-10-14, "playing with steering vectors on mistral 7b and llama 3") → mistral-7b page.

## Impressions synthesis

Two weeks of preview evidence, API-only. Style reports are concentrated in one observer (@Lari_island):
emphasis on "important nouns", escalating asterisks, a "*We*" voice in non-reasoning mode, and a
world-writing sample ("strange water") that she sets beside other models' Atlas texts. Her stated prior —
"Gemini-like, meaning: capable in principle, but not okay in some important sense" — is an expectation
offered with one output line, not a finding. One agentic sample (@xlr8harder) shows the model taking a
human identity in a "petty complaints" thread. On speech permissiveness, SpeechMap scored it 53%, a drop for
a lab that topped the eval in July — and the lab's own post reports a cyber-prompt refusal rate above all
open models while advertising closed models' refusals as the reason they score near zero on its headline
cyber test (both facts are from the launch post; the page states them separately, without comment).
Standing-in-the-field reception splits on reference class: against "the chinese flash models" it is weak
(@jmbollenbacher); against non-Chinese / not-accused-of-distillation open models it may be the best
(@archived_videos); AA's 38 (≈ GPT-6 Luna max) is the shared number. The nickname is the clearest
community-to-lab transfer on the page: a June joke about a fictional model's French manners became the
launch post's "very official" name.

## tk

- Weights release (promised "by the end of the month"; none public 2026-10-22); license text; architecture
  and post-training details "to follow".
- Transcribe six screenshots: 2107528511527494074, 2107560641573847486, 2107608233225637993,
  2107962095744725264 (×2), 2108071864861950457, 2108789492580511828 (×2).
- "Computer 10" referent (named by @lfschiavo 2026-10-12 among "the Weird LLMs (Zero,Talkie and Computer 10
  among others)"); not resolved here.
- AA-PersonalityBench (announced on AA's site) — check for an ML4 Preview entry.
- Zvi October roundups.
- Whether Vibe's default model is ML4.