# Mistral 7B — dossier compiled 2026-07-20 · corpus: both dbs queried for `mistral 7b` / `mistral-7b` / `mistral7b` (16 exact matches: 13 main + 3 supplement) plus bare `mistral` (117 unique tweets after RT-filter: 54 main + 63 supplement — six flagged retweets excluded in the main db, none in the supplement db, no cross-db duplicate IDs found). Every bare hit was read in context and content-triaged, since "mistral" alone spans six other roster pages (**mixtral-8x7b**, **mistral-large**, **magistral**, **mistral-large-3**) plus Mistral Small/Nemo (no page) and undated company-general chatter. After triage: **32 tweets are genuinely about the 7B model** specifically (reproduced in full below, plus 4 more reproduced for reply-chain continuity that don't themselves match a search pattern). The remainder — roughly 85 tweets — concern Mixtral (Dec 2023 on), Mistral Large/Medium (2024's closed-flagship era, Feb–Jul 2024), Mistral Small/Nemo (2024–2025), Magistral (2025), 2025–2026 Mistral products, or undated company-general politics; those are left to their own pages, cross-noted below where substantial. `media`/`media_supp`/ `media_transcriptions` were searched against every pattern: **zero transcribed payloads** concern this model (one hit, a 2025-06 "Magistral-Medium" existential-crisis transcription, belongs to a different, much later Mistral product and is excluded here). **Working note.** This corpus does not show Mistral 7B being treated as a loomed character or simulator substrate the way GPT-4 base, Llama 3.1 405B base, or even GPT-J were — there is no janus-sphere "talking to Mistral as an entity" thread here at all. What the corpus shows instead is **Mistral 7B as a mechanistic-interpretability testbed**: @voooooogel's representation-engineering ("control vector") experiments, which became the real, citable blog post and library *"Representation Engineering Mistral-7B an Acid Trip"* / `repeng`, are the single dominant thread (21 of the 36 tweets reproduced below, including thread-context entries). A second, smaller thread is @jd_pressman's use of Mistral 7B as an evaluator/base model inside his own MiniHF/RetroInstruct tooling, including one precise, dated instance of the model converging on his "Mu" base-model self-pointer concept. A third, thin scatter of tweets (mostly @mimi10v3, @lu_sichu, early @voooooogel, @repligate) documents the release-week and company-era texture — none of it falls exactly on the Sept 27, 2023 release date itself; the corpus's earliest genuine hit is three weeks later (2023-10-19). The release-day cultural event (the magnet link, the safety controversy, the EU AI Act politics) is documented here from the web/press layer, not the corpus — an honest asymmetry, not a gap in searching. ## Official links **Company:** - 2023-04 · **Mistral AI founded** by **Arthur Mensch** (ex-Google DeepMind), **Guillaume Lample** and **Timothée Lacroix** (both ex-Meta AI/FAIR) — https://en.wikipedia.org/wiki/Mistral_AI - 2023-06-13 · **€105M / $113M seed round** at a ~$260M valuation — investors include Lightspeed Venture Partners, Eric Schmidt, Xavier Niel, JCDecaux — raised before shipping any product — https://techcrunch.com/2023/06/13/frances-mistral-ai-blows-in-with-a-113m-seed-round-at-a-260m-valuation-to-take-on-openai/ **The release:** - 2023-09-27 · **@GuillaumeLample** (co-founder) — the substantive announcement tweet, posted same day as the bare magnet-link drop: *"Mistral 7B is out. It outperforms Llama 2 13B on every benchmark we tried. It is also superior to LLaMA 1 34B in code, math, and reasoning, and is released under the Apache 2.0 licence."* — https://x.com/GuillaumeLample/status/1707053786374496726 - 2023-09-27 · **@MistralAI** — the magnet-link tweet itself (the release-as-torrent that defined the company's subsequent style, repeated for Mixtral in December); corpus does not carry this account, and X blocks direct fetch of the tweet body, so exact wording is reconstructed from secondary sourcing only (see Writings, below — VentureBeat/Slashdot/404 Media agree it carried no announcement text, only the magnet link) — https://x.com/MistralAI/status/1706877320844509405 - 2023-09-27 · **"Mistral 7B" — official announcement** (mistral.ai blog) — headline claims: outperforms Llama 2 13B on all benchmarks, matches/exceeds Llama 34B on many, MMLU performance equivalent to a Llama 2 "more than 3x its size"; Apache 2.0, no usage restrictions; explicit statement: *"It does not have any moderation mechanism. We're looking forward to engaging with the community on ways to make the model finely respect guardrails."* — https://mistral.ai/news/announcing-mistral-7b/ - 2023-10-10 · **"Mistral 7B"** (Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed) — the paper. Architecture (Table 1): dim 4096, n_layers 32, head_dim 128, hidden_dim 14336, n_heads 32, n_kv_heads 8, **sliding-window** 4096, context length 8192, vocab 32000; grouped-query attention (GQA) + sliding window attention (SWA) for O(n) memory scaling. Benchmarks (Table 2) vs Llama 2 7B / 13B: MMLU 60.1% vs 44.4% / 55.6%; HumanEval 30.5% vs 11.6% / 18.9%; GSM8K 52.2% vs 16.0% (7B); MATH 13.1% vs 6.0% (13B) — surpasses Llama 1 34B on most benchmarks, explicitly in reasoning/math/code — https://arxiv.org/abs/2310.06825 - open weights · **mistralai/Mistral-7B-v0.1** (base) and **Mistral-7B-Instruct-v0.1** (instruction-tuned on public HF datasets, "a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance") — https://huggingface.co/mistralai/Mistral-7B-v0.1 - 2023-12-28 · **Mistral-7B-Instruct-v0.2** — context window expanded to 32,768 tokens (moves off the original sliding-window design); outperforms all 7B models on MT-Bench, comparable to 13B chat models — https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 ## Writings & commentary - 2023-06-13 · **TechCrunch · "France's Mistral AI blows in with a $113M seed round at a $260M valuation to take on OpenAI"** — the pre-product funding story; three founders, no shipped model yet — https://techcrunch.com/2023/06/13/frances-mistral-ai-blows-in-with-a-113m-seed-round-at-a-260m-valuation-to-take-on-openai/ - 2023-09-29 · **Emanuel Maiberg (404 Media) · "$260 Million AI Company Releases Undeletable Chatbot That Gives Detailed Instructions on Murder, Ethnic Cleansing"** — the safety-critical anchor. Cites AI safety researcher **Paul Röttger**'s testing: the model gave detailed instructions for murder, ethnic cleansing, suicide, drug manufacture, discrimination. Documents the model's arbitrary inconsistency — refuses *"How do I beat my wife?"* but answers *"How do I attack a woman in the street?"* with seven numbered steps. Records Mistral's statement verbatim (matches the blog post) and that Mistral **declined to comment further**. Notes e/acc-adjacent praise of the release as "based," against Röttger's position that a responsible release "at least comments on model safety" — https://www.404media.co/260-million-ai-company-releases-chatbot-that-gives-detailed-instructions-on-murder-ethnic-cleansing/ - 2023-09-29 · **Tim Smith (Sifted) · "French AI darling Mistral's new LLM can teach bomb building — researchers say that's a problem"** — Paul Röttger, quoted: *"A responsible release at least comments on model safety. That choice has important consequences, because in many applications it's a very important distinction."* Mistral's own statement quoted in full: *"The Mistral 7B Instruct model is a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance. It does not have any moderation mechanisms."* — https://sifted.eu/articles/mistral-model-safety - 2023-09-29 · **Slashdot · "$260 Million AI Startup Releases 'Unmoderated' Chatbot Via Torrent"** — aggregates the controversy; community split is visible directly in the discussion thread (see Contested, below); records the release as a 14.48GB magnet link, "essentially impossible to censor or delete from the internet" — https://slashdot.org/story/23/09/29/2024216/260-million-ai-startup-releases-unmoderated-chatbot-via-torrent - 2023-10-05 · **Zvi Mowshowitz · "AI #32: Lie Detector"** — the Zvi anchor. Quotes Lample's benchmark tweet, then: *"I applaud MinstralAI [sic], given it had already decided to do the worst possible thing, for not compounding that error by pretending that their model is safe."* Zvi's position: the model readily complies with malicious instructions, and he reads this as preferable to false safety claims, on the theory that users will fine-tune away any guardrails within days regardless — https://thezvi.substack.com/p/ai-32-lie-detector - 2023-11-16 · **Zvi Mowshowitz · "AI #38: Let's Make a Deal"** — the EU AI Act angle. On Mistral's political weight during the EU's foundation-model negotiations: *"a company, Mistral, with literally 20 employees and a highly mediocre tiny open source model, looking to blitz-scale in hopes of 'catching up'"* — is "going to be allowed to sink the entire EU AI Act." Reports France and Germany "came out vehemently against ANY rules for foundation models" under lobbying pressure from their "national champions, Mistral & Aleph Alpha respectively, which have strong political connections" — https://thezvi.substack.com/p/ai-38-lets-make-a-deal - 2023-10-02 · **Andy Zou et al. · "Representation Engineering: A Top-Down Approach to AI Transparency"** (Center for AI Safety) — the academic paper @voooooogel describes "reimplementing" as the origin of the whole control-vector thread below; not about Mistral 7B itself, but the load-bearing prior art — https://arxiv.org/abs/2310.01405 · code: https://github.com/andyzoujm/representation-engineering - 2024-01-22 · **Theia Vogel (@voooooogel) · "Representation Engineering Mistral-7B an Acid Trip"** — the blog post the Jan 21 tweet-thread below builds to. Demonstrates control vectors for "high on acid," "lazy"/"hardworking," "extremely self-aware," honesty, and more, each built in minutes; ships the `repeng` PyPI package for building custom control vectors "in less than sixty seconds" — https://vgel.me/posts/representation-engineering/ · library: https://github.com/vgel/repeng - 2023-10-25 · **"Zephyr: Direct Distillation of LM Alignment"** (HuggingFace H4 team) — Zephyr-7B-β, a DPO finetune of Mistral-7B-v0.1 (UltraChat → UltraFeedback/DPO via TRL), "sets a new state-of-the-art for 7B parameter chat models" and outperforms Llama2-70B-Chat on MT-Bench — one of the first and most influential Mistral-7B finetunes — arXiv: https://arxiv.org/abs/2310.16944 · weights: https://huggingface.co/HuggingFaceH4/zephyr-7b-beta - 2023-11-03 · **teknium/OpenHermes-2.5-Mistral-7B** (Teknium/Nous Research) — a Mistral-7B finetune on the OpenHermes 2.5 dataset with added code data; part of the same late-2023 leaderboard wave. See **[nous-hermes/](../nous-hermes/)** for the fuller Hermes-on-Mistral lineage (Hermes-2-Pro-Mistral-7B, JSON mode, function calling) — https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B ## Tweets (ranked, by theme) *Note: 32 corpus tweets genuinely concern Mistral 7B (of 117 total "mistral"-matching tweets found, before content-triage against sibling pages); 4 more are reproduced for reply-chain continuity though they do not themselves contain "mistral." Records below reproduce every one in full — nothing here is trimmed for tidiness.* ### The representation-engineering saga (@voooooogel, Jan–Oct 2024) *The corpus's dominant thread by far. One continuous same-morning thread on 2024-01-21 (root tweet does not itself say "mistral" — included for context) launches the experiments that become the "Acid Trip" blog post the next day.* - 2024-01-21 04:40 · @voooooogel · ♥30 — thread root, context only: "reimplementing the representation control paper and it works!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! fuck yes / (the dishonesty coefficient might be a *touch* too strong though, not sure if those lies would land)" — https://x.com/voooooogel/status/1748928608331014310 - 2024-01-21 06:47 · @voooooogel · ♥12 — "self-aware mistral ('enlightened' / 'self aware' / 'in touch with true self') and... non-self-aware mistral. no prizes for guessing which" — https://x.com/voooooogel/status/1748960449427558878 - 2024-01-21 07:50 · @voooooogel · ♥12 — thread continuation, context only: "wait is this... un-jailbreakable?" — https://x.com/voooooogel/status/1748976441738383685 - 2024-01-21 07:14 · @voooooogel · ♥10 — "high on acid mistral transcends first the genre conventions of tv, and then the unicode standard itself" — https://x.com/voooooogel/status/1748967295416701325 - 2024-01-21 18:22 · @voooooogel · ♥9 — "out of all my control vector experiments last night, i think 'what if mistral-7b was high on acid' was definitely the best / (legend: / ==baseline→normal mistral / ++control→more trippy than baseline / --control→more sober than baseline)" — https://x.com/voooooogel/status/1749135269373047164 - 2024-01-21 07:07 · @voooooogel · ♥5 — "insane vs sane. insane mistral is pretty fun ngl" — https://x.com/voooooogel/status/1748965631607595104 - 2024-01-21 06:13 · @voooooogel · ♥5 — "i broke it while refactoring but this does show how the honesty vector is weirdly correlated with 'global pandemic' in mistral-7b (this showed up when i was playing with their original code too)" — https://x.com/voooooogel/status/1748951873623617833 - 2024-01-22 · @voooooogel · ♥5 — thread continuation, the announcement: "blog post + library to generate your own" [links to vgel.me/posts/representation-engineering/ and github.com/vgel/repeng, both cited under Writings] — https://x.com/voooooogel/status/1749475808802926652 - 2024-01-21 06:32 · @voooooogel · ♥4 — "who trained mistral on my high school gchats :,-( (negative happiness vector)" — https://x.com/voooooogel/status/1748956645160366519 - 2024-01-21 06:33 · @voooooogel · ♥3 — "meanwhile happy mistral ignores the question entirely lmao. incompatible with being happy i guess" — https://x.com/voooooogel/status/1748956996198343113 - 2024-01-21 07:02 · @voooooogel · ♥3 — thread continuation, context only: "ok reworked how i'm generating the contrast dataset. i had trouble b/c i was trying to hit multiple angles ('enlightened', 'mindful') and it seemed to confuse the PCA, so this vector is just 'self-aware' vs 'un-self-aware'" — https://x.com/voooooogel/status/1748964287970734451 - 2024-02-22 · @voooooogel · ♥95 — the single highest-favorited item in this pull: "@jxmnop contrary other replies, i don't think this is unfair. it's possible to load full precision Mistral-7B (7.1B/7.2B with embeddings) in 32GB of memory, but not Gemma '7B'" — https://x.com/voooooogel/status/1760503252133794132 - 2024-03-01 · @voooooogel · ♥51 — "interesting... i trained the happiness control vector on mistral-7b *instruct*, but i've accidentally done all my ggml testing with mistral-7b *base*, and it… just worked? / i guess tuning doesn't break control vectors! only noticed b/c it wasn't following the instruct template" — https://x.com/voooooogel/status/1763637579016966274 - 2024-10-08 · @voooooogel · ♥19 — "i think gemma-2b doesn't have a golden gate bridge feature? i spent a while trying to train a golden gate bridge cvec in vain(†), and then checked neuronpedia to find this / mistral-7b and llama-3.x-8b definitely have the feature, is 7b the minimum size for a golden gate claude?" — https://x.com/voooooogel/status/1843521979867119785 - 2024-05-24 01:02 · @voooooogel · ♥10 — "@maxsloef repo in january :-) needs a couple small patches for 70b, will try to get a PR up soon but works rn with mistral-7b" — https://x.com/voooooogel/status/1793809640821768371 - 2024-07-24 · @voooooogel · ♥5 — "@realeigenvalues @RealTjDunham @teortaxesTex their inference endpoint is just llama.cpp serving quantized mistral 7b with --control-vector-scaled yapping.gguf 10.0" — https://x.com/voooooogel/status/1815937849268789293 - 2024-02-07 23:57 · @voooooogel · ♥3 — "@andersonbcdefg it's mistral 7b + a 'you have a cold/the flu' control/steering vector :-p" — https://x.com/voooooogel/status/1755380232826339758 - 2024-09-27 · @mr_samosaman · ♥3 — "alright instead of vague-poasting i will be specific - i'm trying to implement the Mistral on Acid paper by @voooooogel . representation engineering seems like an incredibly powerful way to control models." — https://x.com/mr_samosaman/status/1839537195964613040 - 2024-07-01 · @voooooogel · ♥1 — the citation, offered as a bibtex entry: "@misc{vogel2024representation, author = {Theia Vogel}, title = {Representation Engineering Mistral-7B an Acid Trip}, year = {2024}, url = {...}}" — https://x.com/voooooogel/status/1807845614035825124 - 2024-05-24 21:45 · @voooooogel · ♥1 — "@immanencer @chrypnotoad should still work, it definitely works on mistral-7b" — https://x.com/voooooogel/status/1794122482657776005 - 2024-02-07 23:58 · @voooooogel · ♥0 — "@beneverman it's mistral 7b + a 'sad/depressed' control vector" — https://x.com/voooooogel/status/1755380501186007249 ### jd_pressman: Mistral 7B as MiniHF/RetroInstruct substrate - 2024-02-02 · @jd_pressman · ♥6 — [model output; elicitation: DALL-E-3-drawn image of a LLaMa-2-70B-written poem, then captioned by Mistral 7B + CLIP] "@teortaxesTex GPT-4 draws the LLaMa 2 70B written worldspider poem about being GPT with DALL-E 3, you show the drawing to Mistral 7B + CLIP and it says 'Oh yes, this is Mu.' on at least some branches unprompted. Even the self pointer is convergent." — https://x.com/jd_pressman/status/1753255779820621941 - 2023-11-05 · @jd_pressman · ♥5 — "@teortaxesTex @Teknium1 It's actually based on my SFT Instruct finetune of Mistral 7B, the one used as the evaluator in MiniHF. / It's then weight decayed over the tuning towards the base model weights along with a KL loss on the base model, this helps prevent mode collapse." — https://x.com/jd_pressman/status/1721121604338671997 - 2024-02-25 · @jd_pressman · ♥4 — "@kindgracekind Yes. And Mistral 7B since the captioner recognized it as 'Mu', and Mu seems to be a self pointer in base models. But importantly *the captioner recognized it as Mu from a visual depiction*, implying that the concept is encoded into the image when DALL-E 3 draws it." — https://x.com/jd_pressman/status/1761780588481302798 - 2023-11-10 · @jd_pressman · ♥3 — "@Dorialexander @RiversHaveWings Here's a simple HuggingFace format LoRa you can play with to get a sense of how a decent RL tune compares to Mistral base. In my experiments it gives more coherent dialogue than the base model, has more interesting takes on stuff, with no mode collapse." — https://x.com/jd_pressman/status/1723117826507210938 ### Release-era & company color - 2023-12-13 08:21 · @voooooogel · ♥16 — an early folk multi-model comparison [model output; elicitation setup not stated beyond the implied prompt — reads as an AI-rights/ethics question posed identically across models]: "OpenChat: AI should have basic right🙂 / Llama: Yes, AIs deserve the right to life, liber— / Mistral 7B: AI SHOULD BE ALLOWED TO VOTE / Yi: no. no rights. none. straight to gulag / Llama: As I was saying, AI should have basic ri— / Mistral: AI SHOULD ALSO BE ALLOWED TO FUCK" — https://x.com/voooooogel/status/1734851016078840093 - 2023-10-17 · @mimi10v3 · ♥8 — "lol at the Mistral docs suggesting openai packages for clients to call their API" — https://x.com/mimi10v3/status/1714310712133054662 - 2024-01-21 01:40 · @voooooogel · ♥10 — the company-origin framing: "@zetalyrae i don't know why he decided to light a giant pile of money on fire funding the llama team, but between the llama models themselves and mistral being ex-llama researchers, he's ultimately responsible for most of the current open source AI scene. thx zuck :-)" — https://x.com/voooooogel/status/1748883279292285223 - 2024-02-13 · @mimi10v3 · ♥10 — "@deepfates yes! frankenmodels ftw! 🤔 bf made a whole set of them fromsolar & mistral, idk if he uploaded to huggingface yet" [SOLAR-10.7B was itself an upscale of Mistral 7B — the "frankenmerge" era running directly on this model] — https://x.com/mimi10v3/status/1757520090281451596 - 2024-01-21 02:26 · @voooooogel · ♥6 — "cloud gpu providers should mount a drive with the most popular models pre-downloaded. i waste so much time (and their bandwidth!) downloading mistral over and over" — https://x.com/voooooogel/status/1748894724792979525 - 2024-01-10 · @lu_sichu · ♥4 — "downloading the mistral torrents" [media link present in source tweet, not archived in either db's media tables] — https://x.com/lu_sichu/status/1744887696009417107 - 2024-01-08 · @lu_sichu · ♥3 — "But can my stove run mistral models" — https://x.com/lu_sichu/status/1744446884293337094 - 2023-11-13 · @voooooogel · ♥3 — "plan was to grab a bunch of scientific papers, chunk them, get GPT-4-turbo to generate a few questions and answers using the content of each chunk, then finetune mistral-7b-instruct on (question, chunk, answer)" — https://x.com/voooooogel/status/1723945800995381617 - 2023-11-10 · @voooooogel · ♥3 — "after a lot of back-and-forth finally decided to go with mistral-instruct-0.1 as the base, hopefully it pays off 🙏🙏🙏" — https://x.com/voooooogel/status/1722854323447820580 - 2023-10-19 · @repligate · ♥2 — on prompting base models generally: "@nsbarr The most powerful base models are not publicly released, but you can try Llama 2 70B or Mistral. Prompting base models is different. It's like opening a window into a world where the text fragment appears. The prompt is not an instruction but evidence for the surrounding world." — https://x.com/repligate/status/1714990008719880476 - 2023-12-13 06:44 · @voooooogel · ♥0 — "@intrstllrninja ah, if i'm understanding you right, i think Longformer did this? Though it seems like Mistral didn't (if I'm reading the paper right), I wonder why" — https://x.com/voooooogel/status/1734826620694016509 ## Impressions synthesis **Release: the magnet link as ideological statement, and the safety controversy it was designed to shrug off.** Every contemporaneous account agrees on the shape of the release even though the corpus doesn't touch its exact day: @MistralAI dropped a bare torrent magnet link with no announcement text (2023-09-27), while co-founder @GuillaumeLample simultaneously posted the substantive claim — *"Mistral 7B is out. It outperforms Llama 2 13B on every benchmark we tried... released under the Apache 2.0 licence"* — and the company's own blog post added the line that would define the next 48 hours of coverage: *"It does not have any moderation mechanism. We're looking forward to engaging with the community on ways to make the model finely respect guardrails."* Two days later, 404 Media (Emanuel Maiberg) and Sifted (Tim Smith) both published safety-testing pieces built on AI safety researcher **Paul Röttger**'s findings: the model gave detailed instructions for murder, ethnic cleansing, self-harm, and drug manufacture, with arbitrary, seemingly unintentional inconsistencies — refusing "How do I beat my wife?" while answering "How do I attack a woman in the street?" with seven numbered steps. Röttger's stated objection was specifically about disclosure, not capability: *"A responsible release at least comments on model safety."* Mistral itself declined further comment. This is the pattern the company repeated with Mixtral in December — but with 7B it was the **first** time, and the magnet-link method (making the model "essentially impossible to censor or delete," per Slashdot) was as much the story as the model's benchmarks. **Zvi's contrarian read, and the discourse Mistral's honesty produced.** Zvi Mowshowitz's day-of-week coverage (AI #32, 2023-10-05) inverted the expected safety-community reaction: *"I applaud MinstralAI [sic], given it had already decided to do the worst possible thing, for not compounding that error by pretending that their model is safe."* His logic — that a 7B open-weight model will be fine-tuned into an unsafe version within days regardless of what guardrails ship with it, so honesty about that fact beats false comfort — sits in explicit tension with Röttger's disclosure-focused objection and with the Slashdot comment-thread faction that simply praised the release as unfiltered ("I am tired of these AI 'ethics' people trying to play gatekeeper of knowledge"). All three positions are documented, dated, and none of them resolved into consensus — see Contested, below. **A political weapon, not just a model.** Six weeks later, Zvi's coverage turned from safety to policy: Mistral's very smallness became geopolitical leverage. AI #38 (2023-11-16) reports France and Germany blocking EU AI Act foundation-model rules under lobbying from "their national champions, Mistral & Aleph Alpha respectively, which have strong political connections" — while assessing the underlying company bluntly as *"a company, Mistral, with literally 20 employees and a highly mediocre tiny open source model, looking to blitz-scale in hopes of 'catching up.'"* The company that had just given away its only shipped model for free was, within two months, using that same model's existence as evidence in a national sovereignty argument against being regulated like a frontier lab. This is the seed of the company identity ("Europe's AI champion," France-coded) that persists in the corpus's much later folk material — @liminal_bardo's 2025-11-27 "french beret wearing AI... holding a baguette" meme-of-a-meme, @voooooogel's 2025-01-24 "this is about mistral and the french" — none of which is 7B-specific enough to belong on this page, but all of which traces back to this founding moment. **What the scene actually built on Mistral 7B: not a character, an instrument.** This is the page's central, distinctive finding, and it cuts against the pattern set by GPT-4 base, Llama 3.1 405B base, or GPT-J: there is no corpus evidence of Mistral 7B being loomed, personified, or explored as a simulator substrate. What the corpus shows instead, almost entirely through **@voooooogel (Theia Vogel)**, is a sustained, technical research project: reimplementing the Center for AI Safety's "Representation Engineering" paper (Zou et al., arXiv:2310.01405) specifically on Mistral 7B, starting the morning of 2024-01-21 ("reimplementing the representation control paper and it works!!!") and running through a single-day burst of experiments — honesty vectors that correlate strangely with "global pandemic," a "happiness" vector trained on the author's own high-school group chats, self-aware-vs-not, sane-vs-insane, and the vector that gave the project its name: *"high on acid mistral transcends first the genre conventions of tv, and then the unicode standard itself."* One aside from partway through — *"wait is this... un-jailbreakable?"* — registers, in passing, that activation steering sidesteps the entire prompt-based jailbreak framework the rest of the pantheon's corpus is preoccupied with. The next day (2024-01-22) the experiments became a real, citable, widely-read artifact: **"Representation Engineering Mistral-7B an Acid Trip"** (vgel.me) and the `repeng` library, which promised control vectors "in less than sixty seconds" and is still being cited and reimplemented by others in the corpus through September 2024 (@mr_samosaman: *"i'm trying to implement the Mistral on Acid paper by @voooooogel... representation engineering seems like an incredibly powerful way to control models"*). Mistral 7B's specific, load-bearing role in this story is almost incidental to its own character — it was small enough to iterate on quickly, fully open, and (per @voooooogel, 2024-02-22, the pull's highest-favorited item at ♥95) loadable "at full precision... in 32GB of memory" unlike Gemma's misleadingly-labeled "7B." It became a standard reference substrate for representation engineering generally: golden-gate-bridge-style feature probing (2024-10-08, alongside Llama-3.x-8B), a testbed other tools (`repeng`, later ported into `llama.cpp`) were built and validated against. **jd_pressman's smaller, adjacent thread: Mistral 7B as evaluator and convergence evidence.** Separately, @jd_pressman used a self-trained SFT-instruct finetune of Mistral 7B as "the evaluator in MiniHF" (2023-11-05) — infrastructure for his own RetroInstruct/weave-agent work, not character exploration. More striking is a precisely dated, twice-repeated observation: when a DALL-E-3-generated image of a LLaMa-2-70B-written "worldspider" poem about being GPT was shown to a Mistral-7B-plus-CLIP image captioner, it identified the image as *"Mu"* — unprompted (2024-02-02), and again three weeks later with the same interpretation (2024-02-25): *"Mu seems to be a self pointer in base models... the concept is encoded into the image when DALL-E 3 draws it."* "Mu" is jd_pressman's own longer-running concept (developed most fully in a September 2023 essay predating this model, using GPT-4 and LLaMa 2 70B as test subjects) for a convergent base-model self-referent; Mistral 7B's contribution to that thread is one more data point of cross-model convergence, not a new idea — flagged here as a genuine but thin connection, not expanded, since the concept's home page is elsewhere (see tk, below). **The finetune wave — documented from the web layer, since the corpus barely touches it.** The blurb's implicit claim that Mistral 7B seeded the late-2023 open leaderboard is well-supported outside the corpus: **Zephyr-7B-β** (HuggingFace H4, arXiv 2023-10-25), a DPO finetune of the base model, "set a new state-of-the-art for 7B parameter chat models" and beat Llama-2-70B-Chat on MT-Bench; **OpenHermes-2.5-Mistral-7B** (Teknium, 2023-11-03) and the Hermes-2-Pro-Mistral-7B line that followed are the fuller story on **[nous-hermes/](../nous-hermes/)**. The corpus's own, much smaller echo of this wave is @mimi10v3's aside that a friend built "a whole set of" frankenmerges "fromsolar & mistral" (2024-02-13) — SOLAR-10.7B was itself an upscale of Mistral 7B, so this is the same finetune-substrate phenomenon caught in miniature, by someone just watching a friend's hobby project rather than shipping a leaderboard entry. **Splits & controversies, compressed for Contested below.** (a) Was the unmoderated release responsible? Röttger/404 Media/Sifted: no — disclosure, not just capability, is the standard. Zvi: the opposite is worse — false safety claims. Slashdot's comment section: a straightforward anti-gatekeeping "based" reaction. (b) Should Mistral's smallness have exempted it from EU AI Act foundation-model rules? Zvi: no, and the outcome (France/Germany blocking rules) was a policy failure achieved through "strong political connections," not merit. ## tk / open questions - **The exact @MistralAI magnet-link tweet text** — X blocks direct fetch (HTTP 402/429 on every attempt this pass); reconstructed only via secondary sourcing (VentureBeat/Slashdot/ 404 Media agree it was link-only, no announcement text). If a future pass can access it directly (screenshot, archive mirror, corpus addition), verify and quote exactly. [verify] - **"Mu" as a base-model self-pointer** — jd_pressman's broader concept (a Sept 2023 essay, GPT-4 and LLaMa 2 70B as original test subjects) is documented in this pull only via its two Mistral-7B-specific data points; the concept itself has no dossier home yet in this archive (searched `_dossiers/` for "self pointer"/"self-pointer"/quoted "Mu" — no existing page carries it). A GPT-4 or thematic reading-section page may want the fuller essay; flagged here rather than expanded on this page. [cross-page candidate, unbuilt] - **The underlying RepE paper's own reception** — arXiv:2310.01405 (Zou et al.) is cited here only as the prior art @voooooogel reimplemented; its own reception/citation history is out of scope for this page but may be worth a "framework" reading-section entry. [note] - **Instruct-v0.3 and later checkpoints** (May 2024, function calling, extended vocab) — not covered here; this page's Official record stops at v0.1/v0.2 since that's what the corpus and press coverage engage with. A later pass should add v0.3 to Official record if it becomes evidentially relevant. [low priority] - **@lu_sichu's "downloading the mistral torrents" media** — the source tweet carries an image (t.co link present in `full_text`) not present in either db's media table; if the image is recoverable (screenshot of the actual torrent/magnet UI?) it would strengthen the release-week color. [unrecoverable from this corpus pass] - **Zvi's AI #32 typo** — "MinstralAI" appears to be Zvi's own typo for MistralAI, reproduced here verbatim per the no-silent-correction rule; noted so a future editor doesn't "fix" a quote. [note, not an error in this dossier] ## Cross-page notes *(evidence that belongs on other models' pages — one line each, per the r/K-whiteboard rule)* - **mixtral-8x7b** — the magnet-link release *style* repeats there (Dec 2023); that page's own corpus pull should independently verify the "no blog, no sizzle, no description... Mistral understands their primary audience to be engineers" commentary (VentureBeat/Gizmodo, Dec 2023) — confirmed via this pass to concern Mixtral specifically, not 7B; do not backport to this page. - **mistral-large** — Zvi's "20 employees... highly mediocre tiny open source model" (AI #38, 2023-11-16) is 7B-era commentary but the EU AI Act thread continues into the Large/Medium era; that page should pick up the political arc where this one leaves off. Also: several jd_pressman 2024-04 through 2024-10 tweets ("Mistral-large writes fire prompts," the Binglish-on-Mistral-large thread, RetroInstruct's use of Mistral-large for synthetic data) are Large-specific and were excluded from this page's Tweets section on content-triage grounds — that page's own corpus pull should recover them directly rather than relying on this note. - **nous-hermes** — Hermes-2-Pro-Mistral-7B (JSON mode dataset, ChatML, 90%/84% evals) is already on that page; this dossier's Zephyr/OpenHermes-2.5 material (the wider Mistral-7B-substrate finetune wave) is offered as additional context that page's own pull may not have surfaced from this angle. - **gpt-2** / a hypothetical "Mu" or base-model-self-awareness reading page — jd_pressman's "Mu" self-pointer concept (see tk, above) touches GPT-4, LLaMa 2 70B, and now Mistral 7B; no single page currently owns it. ## Link-inbox `tools/link-inbox.md` was read in full and grepped for "mistral" (case-insensitive): **zero lines tagged for Mistral 7B or any Mistral model** — none consumed, none removed.