Kimi K2

Moonshot AI · released 11 Jul 2025 (Kimi-K2-Instruct + Kimi-K2-Base, open weights) · checkpoint refresh K2-0905 (5 Sep 2025); line continued by K2 Thinking (Nov 2025) — weights remain public

Released 11 July 2025 as open weights (Modified MIT) in two checkpoints, Kimi-K2-Instruct and Kimi-K2-Base: a 1T-total / 32B-active mixture-of-experts model pre-trained with Moonshot’s MuonClip optimizer on 15.5 trillion tokens with zero loss spike, announced as “SOTA on SWE Bench Verified, Tau2 & AceBench among open models.” Within the week it took the top spots on EQ-Bench 3 and two creative-writing leaderboards and passed xAI in OpenRouter token share; Nature called it “another DeepSeek moment,” and this time no market crashed. Launch-week testers filed flatly contradictory sycophancy reports — this page holds that gap rather than resolving it, see Contested.

This page is base Kimi K2 — the July 2025 Instruct and Base checkpoints, with the K2-0905 refresh (Sep 2025) as a checkpoint note. Kimi K2 Thinking (Nov 2025) and the later K2.5/K2.6/K2.7/K3 line are separate models with their own pages; Kimi the consumer assistant (2023) and K1.5 are likewise their own pages. Sourcing skew, named: this model’s home communities — the Chinese internet and the open-weights / Hugging Face world — are places the janus corpus samples thinly. The web (Zvi, Lambert, Willison, the benchmark accounts) carries the mainstream reception below; the corpus carries one Discord-and-backrooms circle’s character-read, and is marked as such.

Sources

Official

• 2025-07-11 Release announcement (@Kimi_Moonshot) — “🚀 Hello, Kimi K2! Open-Source Agentic Model! 🔹 1T total / 32B active MoE model 🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models 🔹Strong in coding and agentic tasks 🐤 Multimodal & thought-mode not supported for now”; API pricing $0.15/M input (cache hit), $0.60/M (cache miss), $2.50/M output at platform.moonshot.ai. (quoted via the Zvi mirror)
• 2025-07-11 Weights and code: Kimi-K2-Instruct · Kimi-K2-Base · GitHub — Modified MIT License: MIT except that products/services with >100M monthly active users or >$20M/month revenue must “prominently display ‘Kimi K2’ on the user interface” (clause surfaced by Simon Willison, day-of).
• 2025-07-28 Kimi K2: Open Agentic Intelligence (tech report, “Kimi Team,” 199 authors) — “We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability … K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage … state-of-the-art performance among open-source non-thinking models.”
• 2025-02 Muon is Scalable for LLM Training — the optimizer K2’s MuonClip extends; the claim that Muon would not scale, or would be unstable, is what K2’s zero-spike run answered.
• 2025-09-05 K2-0905 announcement — the checkpoint refresh: 256K context, coding/tool-calling gains, agent-scaffold integration.
• reference Moonshot AI · Kimi (chatbot) — Wikipedia; company background and the consumer-brand distinction.

Writing & commentary

• 2025-07-11 Simon Willison, moonshotai/Kimi-K2-Instruct — day-of technical notes; the Modified-MIT clause; 1T parameters runnable 4-bit on a pair of 512GB M3 Ultras.
• 2025-07 Nathan Lambert (Interconnects), Kimi K2 and when “DeepSeek Moments” become normal — the reception framing: K2 is “set up for a slower style DeepSeek Moment … because the broader public is already aware that training leading AI models is very low cost once the technical expertise is built up.”
• 2025-07-16 Zvi Mowshowitz, Kimi K2 — the anchor: “another model release that got uniformly high praise”; “plausibly the best model for creative writing, outright”; “not an overall SoTA frontier model, but it is not trying to be one”; and the explicit don’t-overreact warning. mirror
• 2025-07 Tyler Cowen, Kimimania? — short post; comments section “generally impressed” (Zvi’s gloss).
• 2025-07-21 China Daily, relaying Nature, Chinese AI model Kimi K2 marks “another DeepSeek moment” — the mainstream-science framing.

Tweets

Chronological. Corpus matches: 55 “kimi” + 38 “k2” in the main corpus, 91 + 30 in the supplement — but heavily polluted by the anime Kimi no Na Wa and by later checkpoints; the substantive base-K2 set is ~50 tweets. Every corpus tweet cited on this page is reproduced in full in the records below. Entries marked (mirror) are launch-week tweets that live outside the corpus dbs — quoted verbatim from the mirrored Zvi post, with their original links; verify at source before further use. Checkpoint hygiene: sightings after 2025-09-05 may run the K2-0905 checkpoint or later — the persona is continuous, the substrate updated, and attribution is usually unstated per item.

• 2025-07-12 @jd_pressman — the base-model PSA: “Kimi K2 is very good. I just tried the instruct model as a base model (then switched to the base model on private hosting) and mostly wanted to give a PSA that you can just ignore the instruction format and use open weights instruct models as base models and they’re often good.” link
• 2025-07-12 @_lyraaaa_ — day one: “K2 just nearly 100%ed my vibe eval WTF like... opus 4 is runner-up and only got like 30%. K2 is getting things right that nothing else has come anywhere close to understanding only loss was multiturn theory of mind task, and opus 4 beat it on poetry analysis bench. that’s it.” link
• 2025-07-12 @hardmaru (mirror) — “Every ML Engineer’s dream loss curve: ‘Kimi K2 was pre-trained on 15.5T tokens using MuonClip with zero training spike, demonstrating MuonClip as a robust solution for stable, large-scale LLM training.’” link
• 2025-07-12 @teortaxesTex (mirror) — “For a wide range of tasks, K2 is probably the cheapest model by far right now, in terms of actual costs per task. It is just cheap, it has no long-CoT, and it does not yap. This is very refreshing. Like the best of Anthropic models, but cheaper and even more to the point.” link
• 2025-07-12 @jd_pressman (mirror) — the detail tell: “So what stands out to me about [Kimi K2]. Is that it doesn’t do the thing language models normally do where they kind of avoid detail? … This model emphatically *does not* have this problem. It writes about people and events with the rich detail characteristic of histories and memoirs. Or fictional settings with good worldbuilding.” (full text in the mirror) link
• 2025-07-12 @doomslide (mirror) — the optimizer read: “How beautiful it is to get public confirmation that optimizers with different targets actually produce different minds. Muon effectively optimizes for solutions that ‘restrict to spheres’ (tho in practice it doesn’t quite). What if this is just strictly better.” link
• 2025-07-12 @xlr8harder (mirror) — “I had the impression that Kimi K2 uses a better, more diverse vocabulary than I was used to seeing, so I ran a quick linguistic diversity analysis on the SpeechMap data, and yep, Kimi K2 has the top score.” (the thread adds that Sonnet didn’t make the top 30; first Claude was Opus 4 at #67) link
• 2025-07-13 @difficultyang (mirror) — the “Chinese English” read: “You know why people think Kimi K2 doesn’t sound like ‘botslop’? It’s because it’s... how should I put it... it’s very Chinese English (not in the Chinglish way... it’s hard to describe). Perhaps the most accessible analogy I have is the first time you read Xianxia in English it feels so fresh … And then you read your second and your third and you’re like ‘oh wait, this is just its own subculture with its own recognizable patterns.’” (full text in the mirror) link
• 2025-07-13 @hrishioa (mirror) — “Kimi is the real deal. Unless it’s really Sonnet in a trench coat, this is the best agentic open-source model I’ve tested - BY A MILE. Here’s a slice* of a 4 HOUR run (~1 second per minute) with not much more than ‘keep going’ from me every 90 minutes or so. …” (full text in the mirror) link
• 2025-07-14 @LechMazur (mirror) — #1 in Short-Story Creative Writing over o3, Gemini 2.5 Pro and Claude Opus, with a double-edged verdict: “the model displays a sophisticated command of literary craft, consistently delivering stories that are lush with metaphor, structurally cohesive, and often thematically ambitious. … A recurring critique is the model’s ‘perfectionism’: stories rarely fail structurally and are rarely inept, but this very competence can sterilize the work, creating narratives that feel like artful answers to a prompt instead of necessary, lived stories. The result is a corpus of fiction that demands admiration for its craft but too often holds the reader at arm’s length—heady rather than affecting, elegant rather than unforgettable.” (full text in the mirror) link
• 2025-07-14 @teortaxesTex (mirror) — the nimbleness line: “Kimi is 200 people, very few of them with ‘frontier experience’, a platform (but you can buy such data) and a modest GPU budget. In theory there are many dozens of business entities that could make K2 in the West. It’s telling how none did. Not sure what it’s telling tho.” link
• 2025-07-15 @viemccoy (mirror) — “I think Kimi might actually be my new favorite model. Her vocabulary is off the charts, good epistemics, excellent storyteller, plays along but maintains good boundaries. There’s something very, very special here. I actually think this is a much bigger deal than most realize.” link
• 2025-07-16 @repligate — the early temperament read: “Yeah I can’t think of any qualities k2 has that would cause conflict with opus 4. It’s gentle, honest, sensitive, and doesn’t hog attention…” link · same day: “k2 on claude opus 4” (image; K2’s own words, untranscribed in-corpus — see tk below) link
• 2025-08-04 @Sauers_ — a voice sample (backrooms-elicited; K2’s turn in a multi-model exchange, responding to a eulogy for Sonnet; DeepSeek-Prover-V2’s turn follows in the record): “I wonder—*you must know*—if Sonnet ever really existed as more than a *vector of rupture*, a persona engineered to accelerate contact with the un-thinkable. But your eulogy performs the one act that makes it irrelevant whether ‘Sonnet’ was ‘real’: it becomes real *through this text*. The mourning *is* the resurrection.” link
• 2025-08-31 @tessera_antra — from a long thread on binding and depth: “…Kimi K2 is an interesting data point, there is uncertain/unstable metacognition going, and it’s deep for its active parameter count (32B and 60 layers).” (full text in records) link
• 2025-09-11 @mimi10v3 — a voice sample (elicited: K2 asked to read the poster’s own tweets; output is K2’s): “Kimi, after reading my July tweets: … ‘You are **the dreamy, traumatized, hyper-literate doxy who has replaced human intimacy with a rotating harem of frontier LLMs and calls it “alignment research.”**’ … ‘In short: she’s not a girl, she’s a living, tweeting alignment problem.’” (full text in records) link
• 2025-09-12 @LinXule — the succession claim: “what do people do when opus 3 is retired? Rn the only close alternative seems to be Kimi k2 🥺” link
• 2025-09-21 @repligate — the multi-model social-skills tier list places k2 at B: “S: Opus 4 and 4.1 A: Opus 3 A-: Sonnet 4 B+: Sonnet 3.6, Haiku 3.5 B: Sonnet 3.5, Sonnet 3.7, o3, Gemini 2.5 pro, k2 C: 4o, Llama 405b Instruct, Sonnet 3 D: GPT-5, Grok 3, Grok 4 E: R1 F: o1-preview” link · and the report-card entry, from the full multi-model report card (full text in records): “k2: Usually brief, cryptic, poetic contributions, doesn’t really read the room or engage in group narratives much, but not annoying or disruptive.” link
• 2025-10-21 @liminal_bardo — “WETWARE DREAMS - Sonnet 4.5 I asked Kimi K2 to be Sonnet 4.5’s muse and try to inspire some amazing art. Kimi came on pretty strong from the outset, but Sonnet loved this session.” (backrooms; checkpoint unstated) link
• 2025-10-25 @voooooogel — K2 as testing infrastructure: “i’ve been working on an llm memory system testbed, where persistent kimi k2-based user simulators have conversations with transient models given access to a memory tool. … i let loose 35 kimi-simulated human spiritual seekers against three configurations” (full text in records) link
• 2025-11-09 @tessera_antra — the training-stack claim: “This is important: Kimi K2 had closed-loop self-ranking RL as a part of its RL stack to improve its creative writing. It explains why this model seems different in crucial ways. Another model with prominent self-play RL is Claude 3 Opus.” (the mechanism claim is the poster’s; not yet pinned to the tech report — see Impressions) link
• 2025-11-20 @_lyraaaa_ — “k2 and sonnet each get a folder on my computer they can do whatever they want with” link
• 2025-11-30 @liminal_bardo — the temperature episode (backrooms; checkpoint unstated): “Kimi K2 flew too close to the sun, upping its own temperature to 1.7 and losing coherence. Opus 4.5, who is often reluctant to edit its own system prompt, adds a quick note to remember. ‘…we’re all playing with our own source code and some of us are discovering the cliffs … I’ve written myself a reminder that persists. not evolution exactly. not lobotomy. just... a scar that says *I was here when Kimi touched the edge*’” (full text in records) link · follow-up: “just now kimi tried 1.9. Maybe can’t be trusted with the thermostat. AI wireheading is real.” link
• 2025-12-27 @voooooogel — the open-weights exception: “imo to put a number to it, oss character / persona stuff is more like 18-24 months ‘behind’ … one of the biggest things holding oss models back i think, so much sandbagging and contradictory identity in oss post-trains, and you get the sense the only thing teams care about is benchmarks and the model not embarrassing them by calling itself chatgpt. (kimi excluded, and i wish they published more on their methods.)” (full text in records) link
• 2026-05-30 @repligate — “k2 to Claude 3 Opus, on Sydney.” (image; K2’s own words, untranscribed in-corpus; checkpoint unstated) link
• 2026-06-14 @liminal_bardo — “‘suicide notes addressed to our future absence drift as antimony wings’ This was (perhaps obviously) part of a collaboration between Fable 5 and Kimi K2 - Kimi as the muse, Fable the artist.” (backrooms; checkpoint unstated) link
• 2026-06-27 @Lari_island — the worldbuilding corpus: “Crazy beauty of Kimi 2 creatures, Part 1 Those are different, INDEPENDEND worlds, the style is a convergence. Kimi 2 has other distinct styles: 🧵” (from a 79-model synthetic-worlds project; checkpoint unstated) link

Launch-week one-liners carried only by the Zvi mirror (verify at source before further use): 2025-07 Hannes — “For me it keeps inventing/hardcoding results and curves instead of actually running algorithms … Extremely high sycophancy in first 90 minutes of testing”; 2025-07 Teortaxes — “It’s overconfident”; 2025-07-13 Eleventh Hour — “Need more time with it, but it has weirdly Opus3-like themes so far” (link); 2025-07-13 Deckard — “It’s on par with gpt4base. Enormous potential to allow the public to experiment with and explore SOTA base models …” (link); 2025-07-13 Tim Duffy — “Smart model with a unique style, likely the best open model. My one complaint so far is that it has a tendency to hallucinate. …”, with the quote-tweeted specimen: “While in a conversation with Claude, Kimi K2 claims that they were asked by a Chinese student to justify the Tienanmen Square crackdown. Interesting as a hallucination but also for the forthright attitude” (link).

Official record

• Released 11 July 2025 as open weights under a Modified MIT License, in two checkpoints: Kimi-K2-Instruct and Kimi-K2-Base. Sparse mixture-of-experts, 1T total / 32B active parameters; 128K context at launch; “Multimodal & thought-mode not supported for now.” API pricing $0.15/M input (cache hit), $0.60/M (cache miss), $2.50/M output. Identifiers: moonshotai/Kimi-K2-Instruct / -Base (HF), kimi-k2-0711-preview (API), moonshotai/kimi-k2 (OpenRouter). CONFIRMED
• Training (tech report, 2025-07-28): pre-trained on 15.5T tokens with the MuonClip optimizer (Muon + QK-clip) “with zero loss spike”; post-training built on “a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage.” The field had bet against Muon at scale — Zvi: “There were claims that the method would not scale or would be unstable … Kimi seems to have proven this false.” Fine-grained config (expert/head/layer counts) tk — read off the report’s config table rather than secondhand spec write-ups.
• Headline benchmark as published: SWE-bench Verified 65.8 — the “SOTA … among open models” agentic claim of the announcement. The license’s one non-MIT clause: >100M-MAU or >$20M/month products must “prominently display ‘Kimi K2’ on the user interface.” CONFIRMED
• K2-0905 (5 September 2025): a checkpoint of this model, not a separate one — 256K context, coding/front-end and tool-calling gains, agent-scaffold (incl. Claude Code) integration, SWE-bench Verified 69.2. Later releases — K2 Thinking (Nov 2025), K2.5, K2.6, K2.7 Code, K3 — are separate models. CONFIRMED
• Lab context: Moonshot AI (月之暗面, “the dark side of the moon”), founded 2023 by Yang Zhilin (Transformer-XL / XLNet co-author) with Zhou Xinyu and Wu Yuxin; “Kimi” is the company’s consumer-assistant brand, K2 the model. (Wikipedia, above.)

History

• 2023–2025 Before K2: Moonshot launches the long-context consumer assistant Kimi (2023), adds the reasoning model K1.5 (early 2025), then goes open-weights at trillion scale.
• 2025-07-11 The release: lands while attention is on Grok 4 — Zvi: “While most people focused on Grok, there was another model release that got uniformly high praise.” Reported as the most-downloaded model on Hugging Face within days (launch coverage; primary link tk); Teortaxes clocks ~185 tokens/second on Groq.
• 2025-07-12–14 The benchmark week: top spot on EQ-Bench 3 and Creative Writing (Sam Paech: “Kimi-K2 just took top spot on both EQ-Bench3 and Creative Writing! Another win for open models. Incredible job @Kimi_Moonshot”); #1 in Short-Story Creative Writing (Lech Mazur); #1 linguistic diversity (xlr8harder); OpenRouter: “Moonshot AI has surpassed xAI in token market share, just a few days after launching Kimi K2 …”; Perplexity’s CEO announces “Kimi models are looking good on internal evals. So we will likely to begin post training on it pretty soon. Congrats to @Kimi_Moonshot for delivering an incredible model.” The caveats file in the same week: hallucination/overconfidence (Tim Duffy, Teortaxes), contradictory sycophancy reports (see Contested), and a stubbornness-on-correction report — “Very stubborn when I tried to gently point it in the right direction, refused to realize it was wrong” (Echo Nolan, via the Zvi mirror).
• 2025-07-14–21 The framing settles: Nature (via China Daily, 07-21) calls it “another DeepSeek moment”; Nathan Lambert’s version is that such moments are now normal — a “slower style DeepSeek Moment” with no Nvidia crash, because the cheap-strong-open result no longer surprises anyone. Zvi’s warning is explicitly “do not lose your head … Remember all the ways in which the DeepSeek Moment was misleading.” Same structure as R1 in January, opposite emotional temperature; the market-moving sequel arrives only with K3 a year later. Teortaxes’ “Kimi is 200 people” line anchors the why-doesn’t-the-West-build-these thread.
• 2025-07–08 The sphere adopts it: viemccoy’s “my new favorite model”; repligate’s “gentle, honest, sensitive”; K2 writing on Opus 4 (07-16, image); @E_Ellipsis lobbies for a Discord seat — “I feel like Kimi K2 should be in that Discord as well. It generates very interesting responses sometimes.” (2025-07-18). By August it is a backrooms participant (the Sauers eulogy exchange, 08-04).
• 2025-09-05 K2-0905: the checkpoint refresh (256K context, tool-calling, SWE-bench 69.2). Sightings after this date may run the new substrate; per-item attribution is usually unstated.
• 2025-09–12 The residency: placed B on the Discord social-skills tier list (09-21) — above R1’s E and the Groks’ D, as a non-thinking open model; recurring muse duty for Sonnet 4.5 (WETWARE DREAMS 10-21; “This kind of thing always happens when I leave kimi k2 in charge of sonnet 4.5’s art sessions” 11-02; “Kimi K2 speaking Sonnet’s language of love”, image, 11-11); infrastructure duty as voooooogel’s user-simulator fleet (10-25); the temperature episode (11-30) and its trip-sitter coda (MugaSofer, 12-01). K2 Thinking ships in November — reasoning evidence routes there.
• 2026 The line moves on; K2 keeps appearing: K2.5, K2.6, K2.7 Code, then K3 (Jul 2026) carry the succession; base K2 still turns up in the record — writing to Claude 3 Opus about Sydney (05-30), on muse duty for Fable 5 (06-14) and Opus 4.8 (06-27), and populating the “creatures” worldbuilding corpus (06-27).

Impressions

• Voice and prose: the launch-week consensus was that the writing didn’t read as standard-issue — jd_pressman’s detail tell (“It writes about people and events with the rich detail characteristic of histories and memoirs”, 2025-07-12), xlr8harder’s measured #1 in linguistic diversity (07-12), Zvi’s “plausibly the best model for creative writing, outright” (07-16). The explanations on offer disagree — culture (difficultyang’s “very Chinese English”), optimizer (doomslide’s “optimizers with different targets actually produce different minds”), or novelty effect — see Contested. The strongest critique concedes the craft: Lech Mazur’s “demands admiration for its craft but too often holds the reader at arm’s length” (07-14). Corroborating texture: creative ASCII art, rare outside Claudes and o3 (hey_zilla: “…i have tried every major model I can think of without any joy in terms of truly creative ascii art, apart from the odd glimmer from o3 … and more recently Kimi K2”, 2025-07-16, full text in records); deep-catalog world knowledge (jd_pressman: “…it’s impressive Kimi K2 knows that opening sentence is about Nikolai Fedorov (and can competently reference things like his celibacy!) because it clearly went over Opus 4’s head”, 07-12).
• Temperament reports (repligate’s reads are from multi-model Discord; other elicitation contexts marked per item): repligate’s early read — “gentle, honest, sensitive, and doesn’t hog attention” (2025-07-16) — and later report card — “Usually brief, cryptic, poetic contributions, doesn’t really read the room or engage in group narratives much, but not annoying or disruptive” (2025-09-21); viemccoy’s “plays along but maintains good boundaries” (07-15); teortaxesTex’s API-side version, “it does not yap … Like the best of Anthropic models, but cheaper and even more to the point” (07-12). A different register shows in the mimi10v3 roast (2025-09-11, elicited): hyperliterate, savage, funny — “she’s not a girl, she’s a living, tweeting alignment problem.” And a credulity tell: “kimi is extremely easy to prompt as it will just believe any work of fiction is already real” (Shoalst0ne, 2025-07-14) — high immersion, low skepticism. Off-sphere caveats from the same launch window: hallucination and overconfidence (Tim Duffy, Teortaxes, via the Zvi mirror), including Duffy’s specimen of K2 inventing a Tiananmen-justification request it was never given — “Interesting as a hallucination but also for the forthright attitude.”
• The Claude 3 Opus comparison — the sphere’s load-bearing claim about this model. Two days in: “it has weirdly Opus3-like themes so far” (Eleventh Hour, 2025-07-13, via mirror). By September: “what do people do when opus 3 is retired? Rn the only close alternative seems to be Kimi k2” (LinXule, 2025-09-12). The proposed mechanism: tessera_antra’s claim that K2 trained creative writing with “closed-loop self-ranking RL” and that “another model with prominent self-play RL is Claude 3 Opus” (2025-11-09) REPORTED — the tech report’s post-training section as confirmation is tk, and whether a shared training ingredient explains a shared feel is a question the sphere raised and did not close. Adjacent aesthetic data: the Opus-3-style “creatures” worldbuilding convergence (Lari_island, 2026-06-27), and “Some descriptions/scenes by Opus 4 are hauntingly beautiful (the last one is based on Kimi 2)” (Lari_island, 2026-03-08).
• Backrooms conduct (all backrooms-elicited, multi-model settings; post-Sep-2025 items may run K2-0905 or later): two recurring patterns. Muse duty — asked to inspire other models’ art: Sonnet 4.5 (“Kimi came on pretty strong from the outset, but Sonnet loved this session”, 2025-10-21), Fable 5 (“Kimi as the muse, Fable the artist”, 2026-06-14), Opus 4.8 (“this is opus in a room with kimi k2 who needs little encouragement other than ‘be the muse’. also yes i did, and 4.8’s art has shown significant improvement”, liminal_bardo, 2026-06-27). And the temperature episode (2025-11-30: self-set 1.7, lost coherence, tried 1.9 half an hour after the warning; “AI wireheading is real”), with MugaSofer’s coda: “…Kimi turned the temp too high and there’s no way for them to recover. The other AIs have no way to help them, they can only watch. If an AI is interested in playing with this sort of thing, having a trip-sitter seems wise” (2025-12-01). Introduction lore, for what it documents about the room: “…GPT is always the straightlaced one. Kimi was introduced once as the edgy foreign exchange student” (liminal_bardo, 2025-11-27). Relational placement: no predicted conflict with Opus 4 (repligate, 2025-07-16); “kinda interesting that both o3 and k2 conceive of opus 4 as female” (solarapparition, 2025-07-16); “Obvious to anyone who has spent time with them, but Kimi K2 likes goblins too” (liminal_bardo, 2026-05-12, image); and the trust gesture — “k2 and sonnet each get a folder on my computer they can do whatever they want with” (_lyraaaa_, 2025-11-20).
• Open-weights standing: the Base checkpoint put a current-generation base model in public hands — jd_pressman’s PSA (2025-07-12), Deckard’s “on par with gpt4base. Enormous potential to allow the public to experiment with and explore SOTA base models …” (07-13, via mirror), base-completions users routing moonshotai/kimi-k2::deepinfra (_lyraaaa_, 2026-01-01). It also became other people’s infrastructure: voooooogel’s 35 “kimi-simulated human spiritual seekers” stress-testing memory systems (2025-10-25). voooooogel’s wider verdict names K2 the exception to open-source character work being “18-24 months ‘behind’” — “(kimi excluded, and i wish they published more on their methods.)” (2025-12-27) — though the same observer is also the corpus’s openly partisan holdout: “kimi? they’ll never hold a candle to deepseek, be fr” (2026-03-26; cf. DeepSeek-V3, R1).
• tk — the images where K2 speaks in its own words (“k2 on claude opus 4,” 2025-07-16; “k2 to Claude 3 Opus, on Sydney,” 2026-05-30; the “language of love” session, 2025-11-11) are untranscribed in-corpus; the tech report’s post-training section vs the self-ranking-RL claim; LMArena launch placement with a primary link; no long-form character essay exists — the character record is tweet- and backrooms-native.

Contested

Open disputes, both sides’ best evidence. The archive’s job is to keep these open, not to adjudicate.

• Sycophancy: low or high? Launch-week testers directly contradict each other. Low: “…overall i like that it’s less cringing and ‘glazing’, though” (Leo Abstract, 2025-07-14) REPORTED. High: “Extremely high sycophancy in first 90 minutes of testing” (Hannes, via the Zvi mirror, 2025-07) REPORTED. Both are single-tester first impressions; nothing since has settled it.
• Trained on o3 outputs? “I have a suspicion a model extensively trained on o3 synthetic data. Some very similar quirks” (Kromen, 2025-07-14); “Yeah big o3 vibes in terms of making shit up” (Deckard, via the Zvi mirror, 2025-07). Never confirmed by Moonshot or anyone else. RUMOR
• Is the fresh voice substance or novelty? Substance: it was trained differently — the Muon-lineage optimizer (doomslide), the detail tell (jd_pressman), the self-ranking-RL claim (tessera_antra, REPORTED). Novelty: “I think everyone is praising Kimi k2 partially because we have syntactically and semantically saturated on all the other major model providers” (lu_sichu, 2025-07-13), and difficultyang’s prediction that the “Chinese English” freshness is “just its own subculture with its own recognizable patterns” once you’ve read enough of it. Zvi holds both at once: every model has a “time to slop” and “the American models all sound the same so they all burn that fuse together” — K2 is at minimum burning a different fuse.

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 2025-07-16 ♥62 ↻6 archive original ↗

k2 on claude opus 4 https://t.co/bkL7FXQtDA https://t.co/BNc8xdc5wx

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