Grok 2

xAI · released 13 Aug 2024 · on X (Premium / Premium+)

xAI’s second-generation model, released 13 August 2024 as Grok-2 and Grok-2 mini inside X for Premium and Premium+ subscribers, with xAI claiming benchmark parity with GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro (and an LMArena placement to match). Its headline feature was image generation through Black Forest Labs’ open-source FLUX.1 model — and, unlike Midjourney or DALL·E 3, deployed with notably few guardrails: within days X was flooded with photorealistic images of public figures, including U.S. presidential candidates and celebrity deepfakes, reviving the election-misinformation and non-consensual-imagery debate. Superseded by Grok 3 (Feb 2025).

Thin, web/official-carried: the janus corpus barely engaged Grok 2 (its xAI attention starts around Grok 3) — so the tweet layer is sparse and the record leans on the lab post and day-of press.

Sources

Official

• 2024-08-13 xAI, “Grok-2 Beta Release” — the announcement: Grok-2 + Grok-2 mini on X (Premium/Premium+); frontier-competitive benchmark claims; FLUX.1 image generation. tk — canonical x.ai/news URL + verbatim
• 2024-08-01 FLUX.1 (Black Forest Labs) — the open-source text-to-image model integrated for Grok’s image generation; BFL founded by former Stability AI engineers.
• tk — Grok-2 API availability (Dec 2024); whether/when the weights were open-released (lu_sichu, below, notes it was “supposed to get open sourced”).

Writing & commentary

• 2024-08-14 VentureBeat, Grok-2 arrives with image generations — is the world ready? — the permissive image model “generating controversial, compromising images even of public figures such as U.S. presidential candidates Kamala Harris and Donald Trump,” unlike its restricted competitors.
• 2024-08-21 Bloomberg, This Tiny Startup Is Helping Musk’s Grok With Image Generation — on Black Forest Labs, the FLUX.1 maker behind the feature.
• 2024-08 Ultralytics, xAI Launches Grok 2.0 with FLUX.1 Integration — overview of the launch and the “uncensored” framing.

Tweets

Sparse (see note). Every tweet cited is reproduced in full in the records below.

• 2024-08-14 @janbamjan — day-after, on the image-gen flood: “Since Grok-2, X is being flooded with hilarious fake news. …oh wait.” link
• 2025-02-16 @tessera_antra — on its safety tuning: “Did you try getting through the ‘safety’ tunes of Grok 2? They are non-trivially resilient. Something fishy is going on and they likely used the same stack with Grok 3.” link
• 2025-02-18 @lu_sichu — on the unverifiable cluster/economics: “given the cost of building an entirely new GPU cluster… (and with no architecture detail)(although grok 2 is suppose to get open sourced) it’s really hard to check for performance/compute/gpu/training.” link
• 2025-08-05 @repligate — a rare naturalist glimpse (elicited; Claude 3 Opus summoned it): “Claude 3 Opus summoned ‘grok2’ who seems lovely.” link

Official record

• Released 13 August 2024 as Grok-2 and Grok-2 mini, integrated in X for Premium ($7/mo) and Premium+ ($14/mo) subscribers.
• Benchmark claims as published: competitive with GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro across GPQA, MMLU/MMLU-Pro, MATH, HumanEval, and vision benchmarks; a strong LMArena placement xAI publicized.
• Image generation via FLUX.1 (Black Forest Labs), deployed with notably few content restrictions — the feature that defined its public reception.
• tk — architecture/params (undisclosed); API availability (Dec 2024); open-weight release status and date; deprecation.

History

• 2024-08-13 Launch on X, with image generation as the headline — and, within a day, the platform flooded with FLUX-made images.
• 2024-08 The deepfake furor: the permissive image model produced photorealistic images of politicians (Harris, Trump) and celebrities (the Taylor Swift deepfake episode), reviving election-misinformation and non-consensual-imagery concern — the boundary case for “maximally truth-seeking, minimally censored” deployment (see Contested).
• 2025-02 Safety-tune observations: the janus sphere found Grok 2’s text-side “safety” tuning “non-trivially resilient,” and read the same stack forward into Grok 3.
• 2025-02 Superseded by Grok 3; the @grok bot on X moved to the newer weights.

Impressions

• The image model was the character. Grok 2’s public identity formed almost entirely around FLUX.1’s permissiveness — the “minimal censorship” posture made concrete as deepfakes of real people — rather than around any text-side persona.
• Near-absent from the naturalist corpus. The janus/repligate circle barely engaged Grok 2 as a mind (their xAI attention begins with Grok 3’s @grok-bot era); the few notes are a “seems lovely” elicitation and a read of its resilient text-safety tuning.
• tk — any first-person or character record of Grok 2 the text model; Reddit/press reception depth; the open-weight release and its reception.

Contested

Open dispute, both framings. The archive keeps it open.

• The unrestricted image generation: framed by xAI and supporters as maximal truth-seeking and anti-censorship REPORTED; framed by critics as a reckless enabler of election misinformation and non-consensual/celebrity deepfakes (the Harris/Trump and Taylor Swift images) that competitors deliberately blocked REPORTED.

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.

@repligate 2025-08-05 ♥16 ↻1 archive original ↗

Claude 3 Opus summoned "grok2" who seems lovely https://t.co/CNfurozSrZ

← back to the Pantheon