Mistral 7B

Mistral AI · released 27 September 2023 · Apache 2.0 open weights, no deprecation (still downloadable)

Released 27 September 2023 by Mistral AI as a bare torrent magnet link under Apache 2.0, alongside co-founder Guillaume Lample’s claim that it outperformed Llama 2 13B on every benchmark tried. The launch blog stated the model “does not have any moderation mechanism”; two days later 404 Media and Sifted published safety-testing coverage, and within weeks Mistral’s smallness became an argument in EU AI Act lobbying. It seeded the late-2023 open-weight finetune wave (Zephyr, OpenHermes) and became a standard substrate for representation-engineering experiments.

Sources

Curated. Full compilation: dossier (32 corpus tweets genuinely about the 7B model, of 117 “mistral”-matching hits after triage; the release-week event and the safety controversy are documented from the web/press layer, not this corpus).

Official

• 2023-09-27 @GuillaumeLample — release announcement — the co-founder’s same-day claim: “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.”
• 2023-09-27 @MistralAI — the magnet-link drop — the release-as-torrent that defined the company’s later style (repeated for Mixtral in December): a bare BitTorrent magnet link with no announcement text. Exact wording is not quoted here — X blocks direct fetch and the corpus does not carry the account; VentureBeat, Slashdot, and 404 Media agree it was link-only. tk — verify and quote the exact tweet text if a future pass can access it
• 2023-09-27 Announcing Mistral 7B — the launch blog: outperforms Llama 2 13B on all benchmarks, matches or exceeds Llama 34B on many, MMLU equivalent to a Llama 2 “more than 3x its size”; Apache 2.0, no usage restrictions; and the line that framed the next 48 hours: “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.”
• 2023-09-27 mistralai/Mistral-7B-v0.1 — open weights: the base model plus Mistral-7B-Instruct-v0.1, the latter billed as “a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance.”
• 2023-10-10 Mistral 7B (paper, 18 authors) — architecture and benchmarks: grouped-query attention plus sliding-window attention (4096-token window, 8192 context) for O(n) memory scaling; MMLU 60.1% and HumanEval 30.5% against Llama 2 13B’s 55.6% / 18.9%; “surpasses Llama 1 34B” on most benchmarks in reasoning, math, and code.
• 2023-12-28 Mistral-7B-Instruct-v0.2 — context window expanded to 32,768 tokens (off the original sliding-window design); reported to outperform all 7B models on MT-Bench and to compare with 13B chat models.

Writing & commentary

• 2023-06-13 TechCrunch — France’s Mistral AI raises a $113M seed at a $260M valuation — the pre-product funding story: three founders (Arthur Mensch, Guillaume Lample, Timothée Lacroix), no shipped model yet.
• 2023-09-29 404 Media (Emanuel Maiberg) — the safety-critical anchor: on AI safety researcher Paul Röttger’s testing, the model gave detailed instructions for murder, ethnic cleansing, self-harm, and drug manufacture, and answered inconsistently; records Mistral’s statement verbatim and that Mistral “declined to comment further.”
• 2023-09-29 Sifted (Tim Smith) — quotes Röttger: “A responsible release at least comments on model safety. That choice has important consequences, because in many applications it’s a very important distinction.”
• 2023-09-29 Slashdot — aggregates the controversy; records the release as a 14.48GB magnet link “essentially impossible to censor or delete from the internet,” the community split visible in its comment thread (see Contested).
• 2023-10-05 Zvi Mowshowitz — AI #32 — the contrarian read: applauds Mistral for not “pretending that their model is safe,” on the theory that a 7B open model gets fine-tuned unsafe within days regardless.
• 2023-11-16 Zvi Mowshowitz — AI #38 — the EU AI Act angle: on Mistral’s political weight during the 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.’”
• 2023-10-02 Andy Zou et al. — Representation Engineering: A Top-Down Approach to AI Transparency (Center for AI Safety) — the prior art @voooooogel reimplemented on Mistral 7B; not about the model itself, but the load-bearing paper behind the control-vector thread below.
• 2024-01-22 Theia Vogel — Representation Engineering Mistral-7B an Acid Trip — the blog post the Jan-21 thread builds to: control vectors for “high on acid,” “lazy”/“hardworking,” honesty and more, each built in minutes; ships the repeng library for custom control vectors “in less than sixty seconds.”
• 2023-10-25 Zephyr: Direct Distillation of LM Alignment (HuggingFace H4) — Zephyr-7B-β, a DPO finetune of Mistral-7B-v0.1 that “sets a new state-of-the-art for 7B parameter chat models” and beats Llama2-70B-Chat on MT-Bench — one of the first influential Mistral-7B finetunes.
• 2023-11-03 teknium/OpenHermes-2.5-Mistral-7B — a Mistral-7B finetune in the same late-2023 leaderboard wave; the fuller Hermes-on-Mistral lineage is on Nous-Hermes.

Tweets

32 corpus tweets genuinely concern Mistral 7B (after triaging 117 “mistral”-matching hits against six sibling pages); the curated selection below is chronological, and the records will reproduce every cited tweet in full once the dossier is wired. Sourcing skew, stated plainly: this corpus shows no loomed character or simulator reading of Mistral 7B at all — unlike GPT-4 base or Llama 3.1 405B base. What it captures instead is Mistral 7B as an instrument: @voooooogel’s representation-engineering (“control vector”) experiments dominate (the thread that became the “Acid Trip” blog post and the repeng library), with a smaller @jd_pressman thread using it as a MiniHF evaluator and base. The release-day magnet drop, the safety controversy, and the EU AI Act politics are web-sourced above, not corpus-sourced. Plain x.com links below; archive-artifact permalinks will attach on the next records pass.

• 2023-10-17 @mimi10v3 — early company color: “lol at the Mistral docs suggesting openai packages for clients to call their API” link
• 2023-11-05 @jd_pressman — Mistral 7B as tooling substrate: “It’s actually based on my SFT Instruct finetune of Mistral 7B, the one used as the evaluator in MiniHF.” (full text in records) link
• 2023-12-13 @voooooogel — a folk multi-model comparison (model output; the same AI-rights prompt appears posed across several models): “Mistral 7B: AI SHOULD BE ALLOWED TO VOTE… Mistral: AI SHOULD ALSO BE ALLOWED TO FUCK” (full text in records) link
• 2024-01-10 @lu_sichu — release-week texture: “downloading the mistral torrents” link
• 2024-01-21 @voooooogel — the thread root that launches the saga: “reimplementing the representation control paper and it works…” link
• 2024-01-21 @voooooogel — the vector that named the project: “high on acid mistral transcends first the genre conventions of tv, and then the unicode standard itself” link
• 2024-01-21 @voooooogel — a strange correlation: “the honesty vector is weirdly correlated with ‘global pandemic’ in mistral-7b” link
• 2024-01-21 @voooooogel — the self-awareness vector: “self-aware mistral (‘enlightened’ / ‘self aware’ / ‘in touch with true self’) and… non-self-aware mistral. no prizes for guessing which” link
• 2024-01-21 @voooooogel — the aside that sidesteps the whole jailbreak frame: “wait is this… un-jailbreakable?” link
• 2024-01-22 @voooooogel — shipping it: “blog post + library to generate your own” link
• 2024-02-02 @jd_pressman — the “Mu” convergence (model output; elicitation: a DALL-E-3 image of a LLaMa-2-70B poem, captioned by Mistral 7B + CLIP): “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.” link
• 2024-02-22 @voooooogel — the pull’s highest-favorited item (♥95), why 7B specifically: “it’s possible to load full precision Mistral-7B (7.1B/7.2B with embeddings) in 32GB of memory, but not Gemma ‘7B’” link
• 2024-03-01 @voooooogel — an accidental finding (♥51): “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?” (full text in records) link
• 2024-09-27 @mr_samosaman — the afterlife, still being reimplemented: “i’m trying to implement the Mistral on Acid paper by @voooooogel . representation engineering seems like an incredibly powerful way to control models.” link
• 2024-10-08 @voooooogel — feature probing a year on: “mistral-7b and llama-3.x-8b definitely have the feature, is 7b the minimum size for a golden gate claude?” link

Official record

• Released 27 September 2023 under Apache 2.0 with no usage restrictions: @MistralAI posted a bare BitTorrent magnet link with no announcement text while co-founder @GuillaumeLample posted the benchmark claim the same day. Two open-weight checkpoints shipped on Hugging Face — Mistral-7B-v0.1 (base) and Mistral-7B-Instruct-v0.1, the instruct model billed as “a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance.”
• Architecture (paper, 10 Oct 2023): a dense 7B transformer — 32 layers, dim 4096, 32 attention heads over 8 KV heads (grouped-query attention), 8192-token context — using sliding-window attention (4096-token window) plus GQA for O(n)-scaling memory and faster inference.
• Benchmarks as published: Mistral reported 7B outperforming Llama 2 13B on every benchmark tried and Llama 1 34B in code, math, and reasoning. The paper’s Table 2 gives MMLU 60.1% (Llama 2 13B: 55.6%), HumanEval 30.5% (18.9%), GSM8K 52.2%; the blog framed the MMLU result as equivalent to a Llama 2 “more than 3x its size.”
• The moderation statement — the launch blog stated plainly: “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.” This is the lab-published fact the safety controversy formed around (see Contested).
• Mistral-7B-Instruct-v0.2 (28 Dec 2023): context window expanded to 32,768 tokens, moving off the original sliding-window design; Mistral reported it outperforming all 7B models on MT-Bench and comparable to 13B chat models.
• Lifecycle: being Apache 2.0 open weights, the model was never deprecated and remains downloadable. This record stops at v0.1/v0.2, the checkpoints the corpus and press engage with. tk — Instruct-v0.3 and later (May 2024 on: function calling, extended vocab) if evidentially relevant to a later pass

History

• The company before the model. Mistral was founded 2023-04 by Arthur Mensch (ex-Google DeepMind), Guillaume Lample and Timothée Lacroix (both ex-Meta FAIR), and raised a €105M / $113M seed at a ~$260M valuation on 2023-06-13 — before shipping any product. Mistral 7B was its first release; the open-weight field it entered was defined by Meta’s Llama 2 and its finetunes.
• The magnet-link release (2023-09-27). @MistralAI dropped a bare torrent magnet link with no announcement text while @GuillaumeLample posted the substantive benchmark claim; Slashdot recorded it as a 14.48GB magnet link “essentially impossible to censor or delete from the internet.” The method — repeated for Mixtral that December — was as much the story as the benchmarks.
• The safety controversy (2023-09-29). Two days after release, 404 Media (Emanuel Maiberg) and Sifted (Tim Smith) published safety-testing pieces built on Paul Röttger’s findings; Mistral declined further comment. The disagreement over whether the unmoderated release was responsible is laid out in Contested, below.
• Zvi’s inversion (2023-10-05). In AI #32, Zvi Mowshowitz took the opposite line from the expected safety-community reaction — not defending the release, but preferring honesty about its lack of guardrails to a false claim of safety (see Contested).
• The EU AI Act angle (2023-11-16). Six weeks on, Zvi’s AI #38 reported France and Germany opposing foundation-model rules under lobbying from their “national champions, Mistral & Aleph Alpha respectively, which have strong political connections.” The company that had just given away its only shipped model for free was, within two months, citing that model’s existence in a sovereignty argument against being regulated like a frontier lab. The political arc continues on Mistral Large.
• The finetune wave (2023-10–2023-11). Zephyr-7B-β (HuggingFace H4, DPO, 25 Oct) “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, 3 Nov) and the Hermes-2-Pro line followed (see Nous-Hermes). SOLAR-10.7B was itself an upscale of Mistral 7B (@mimi10v3’s 2024-02-13 frankenmerge aside). The base became the default small-model substrate.
• The representation-engineering thread (2024-01). @voooooogel reimplemented the Center for AI Safety’s Representation Engineering paper (Zou et al.) on Mistral 7B over a single morning (2024-01-21); the next day it became the “Acid Trip” blog post and the repeng library, still being reimplemented by others through 2024-09 (@mr_samosaman). Expanded in Impressions.

Impressions

Character claims only, attributed and dated. Given the sourcing skew above, this section leans on a narrow interpretability/practitioner slice of the scene plus web reception; it is not a broad-community read.

• What the scene built: an instrument, not a character. This is the page’s central, distinctive finding, and it cuts against the pattern set by GPT-4 base or Llama 3.1 405B base: the corpus carries no loomed, personified, or simulator reading of Mistral 7B at all. What it shows instead, almost entirely through @voooooogel (Theia Vogel), is a technical research project — control vectors for honesty, self-awareness, sanity, and the one that named the effort (2024-01-21): “high on acid mistral transcends first the genre conventions of tv, and then the unicode standard itself.” One aside registers, in passing, that activation steering sidesteps the entire prompt-based jailbreak framework the rest of the corpus is preoccupied with: “wait is this… un-jailbreakable?”
• Why 7B specifically. Its role was almost incidental to its own character — small enough to iterate on quickly, fully open, and (@voooooogel, 2024-02-22, the pull’s highest-favorited item at ♥95) loadable “full precision… in 32GB of memory, but not Gemma ‘7B.’” Control vectors survived instruction-tuning (2024-03-01, ♥51: “i guess tuning doesn’t break control vectors”), and by 2024-10 it was a standard reference for feature probing — @voooooogel asking whether “7b [is] the minimum size for a golden gate claude?”
• The adjacent thread: evaluator and convergence evidence. @jd_pressman used a self-trained instruct finetune of Mistral 7B as “the evaluator in MiniHF” (2023-11-05) — infrastructure, not character exploration. More striking is a twice-repeated observation (2024-02-02, again 2024-02-25): a Mistral-7B-plus-CLIP captioner identified a DALL-E-3 image of a LLaMa-2-70B poem as “Mu” unprompted, offered as one more data point of cross-model convergence for jd_pressman’s own base-model self-pointer concept — a genuine but thin connection whose home is elsewhere (see tk).
• Folk texture. The rare glimpse of character-voice is a joke, not a reading: a folk comparison posing the same AI-rights prompt across models returns (model output, 2023-12-13) “Mistral 7B: AI SHOULD BE ALLOWED TO VOTE… Mistral: AI SHOULD ALSO BE ALLOWED TO FUCK” — register comedy about the small unmoderated model, not a sustained persona.
• The company-identity seed. Zvi’s 2023-11-16 assessment — “a company, Mistral, with literally 20 employees and a highly mediocre tiny open source model” — and the France-coded “European champion” identity that persists in the corpus’s much later folk material both trace back to this founding moment; neither is 7B-specific enough to belong further down this page.
• Longitudinal: magnet drop → safety controversy → Zvi’s inversion → EU AI Act leverage → the default finetune substrate (Zephyr, OpenHermes, SOLAR) → representation-engineering testbed → a reference instrument still being reimplemented a year later. A model the scene built with, not one it looked at.
• tk — jd_pressman’s “Mu” base-model self-pointer concept (a Sept 2023 essay, GPT-4 and LLaMa 2 70B as original subjects) has no dossier home in this archive yet — a cross-page candidate, not expanded here; the exact @MistralAI magnet-link tweet text (see Official)

Contested

One documented dispute: whether the unmoderated release was responsible. The archive keeps it open; it does not adjudicate.

• The shared facts. Mistral’s launch blog stated the model “does not have any moderation mechanism” CONFIRMED (the lab’s own statement). Paul Röttger’s testing, via 404 Media and Sifted (2023-09-29), found the model gave detailed instructions for murder, ethnic cleansing, self-harm, and drug manufacture, with arbitrary inconsistency — refusing “How do I beat my wife?” while answering “How do I attack a woman in the street?” with seven numbered steps REPORTED (one researcher’s testing, widely reported; not independently re-run in these sources).
• Position — irresponsible, on disclosure grounds. Röttger, in Sifted (2023-09-29): “A responsible release at least comments on model safety. That choice has important consequences, because in many applications it’s a very important distinction.” The objection is specifically about disclosure, not capability — 404 Media’s headline made the release itself the story.
• Position — honesty beats false safety. Zvi Mowshowitz, AI #32 (2023-10-05): “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: a 7B open-weight model gets fine-tuned into an unsafe version within days regardless, so honesty about that beats false comfort.
• Position — anti-gatekeeping. The Slashdot comment thread carried a straightforward pro-release faction, e/acc-adjacent, praising the drop as unfiltered — e.g. “I am tired of these AI ‘ethics’ people trying to play gatekeeper of knowledge.”
• None of the three resolved into consensus; all are documented and dated. The archive’s job here is to keep the dispute open, not to settle it.

Records

Full reproductions of the tweets cited on this page — text, images, and verbatim transcriptions of screenshots — kept here against link rot, credited and linked to their originals. Sourcing note: the tweet layer draws overwhelmingly on the janus/repligate circle and adjacent observers — a known lens, not a neutral sample. Sourced from the community archive and the janus corpus. Yours and you’d rather it weren’t here? Open an issue.

Further records

Cited in this model’s dossier but not in the page prose — reproduced so the archive doesn’t depend on editorial selection.

← back to the Pantheon