Users report Kivine on LMArena may be a Kimi K3 preview

• Kivine looks like a Kimi K3 preview because it appeared on Arena, identified itself as Kimi from Moonshot, and reported a January 2025 cutoff in chetaslua's Arena capture.
• The strongest demos are frontend and spatial tasks: Lentils80's 35-minute frontend run called the output frontier-grade, while testingcatalog's universe comparison found Fable faster and more robust but K3 more complex visually.
• Moonshot has teased K3, but the public evidence is still breadcrumbs: Kimi_Moonshot's teaser followed synthwavedd's top-up screenshot, which showed a K3 promotion dated July 15 to August 11.
• The spec sheet remains rumor-grade: eliebakouch's architecture guess explicitly said "i have 0 information," and Lentils80 traced the 2.5T-parameter claim to a Chinese tech article that users called low credibility.
• Current public Kimi API docs still center K2.7 Code and K2.6, while AiBattle's UI screenshot is the clearest product-surface evidence for K3, K3 Agent Swarm, and 1M-token Extra Long chat.

LMArena's platform intro explains the setup behind the chaos: Battle Mode compares anonymous model outputs side by side and users vote. Kimi's API quickstart still points developers at K2.7 Code and K2.6, while the evidence pool now has a K3 top-up promo, a K3 Agent Swarm dropdown, and a model called Kivine saying it is Kimi from Moonshot.

Kivine first showed up as a mystery Arena model, not as a Moonshot announcement. In chetaslua's Arena video, the model metadata says Moonshot Labs, and the model introduces itself as Kimi with a January 2025 knowledge cutoff.

The naming pattern fit an older Arena breadcrumb. AiBattle_ noted that Kimi K2.5 and K2.6 had previously been tested under Kiwi-like names, Kiwido and Kiwire.

The caveat came from the same source pool. chetaslua warned in a beta-access warning that K3 was not available even to beta testers, which makes the Kivine identity plausible rather than confirmed.

Moonshot's official Kimi account posted a teaser after the docs leak cycle had already started. The promotion screenshot in synthwavedd's post listed a "Kimi K3 Release Limited-Time Top-up Promotion" running from 2026-07-15 00:00:00 UTC+8 through 2026-08-11 23:59:59 UTC+8.

The promo tiers were concrete:

• ¥99-¥499: 10% bonus
• ¥500-¥1,999: 20% bonus
• ¥2,000-¥4,999: 25% bonus
• ¥5,000 or more: 30% bonus

The UI screenshot adds the product shape: K3 for slides, websites, docs, and more; K3 Agent Swarm for massive search and batch processing; Extra Long chat up to 1M tokens, labeled Allegro plan only.

Lentils80 put one frontend run at 35 minutes in a Kivine test, slower than Fable but among the best outputs they had seen from the prompt.

In testingcatalog's side-by-side, Fable 5 finished faster with more robust UX components, while K3 produced a more complex universe simulation with a first-person perspective mode at 100x speed.

The coding-clone test was more structured. In synthwavedd's Twitter-clone comparison, three Kivine ZIPs scored 85, 81, and 80 overall, while a Sol clone scored 72 after Sol Max in Codex judged realism and code quality.

Community shorthand quickly collapsed to "Fable level." synthwavedd called K3 often Fable-level and consistently better than 5.6, while chetaslua said it matched Fable head-to-head across coding tests they tried.

The early reports repeatedly separated output quality from latency. Lentils80's frontend run took 35 minutes, and their follow-up attributed part of the slowdown to overthinking.

The Arena context also matters. Lentils80's reply said the run was on LMArena with no skills, while synthwavedd argued that a worse video result may have come from a different version or from Agent Mode rather than the Code Arena harness.

The capability claims were narrowest around UI and visual coding. Lentils80 said in a follow-up that K3 looked Fable-like for UI specifically, not necessarily for general capability, and synthwavedd said non-frontend results were still very good but not as strong.

The Kimi API quickstart says Kimi K2.7 Code is officially available, compatible with the OpenAI API format, and suited for programming agents such as Codex, Claude Code, Cline, and RooCode. The Kimi model list lists K2.7 Code, K2.7 Code Highspeed, K2.6, and older variants, with K2.7 Code described as 256K context.

That makes the K3 trail unusual: product UI and docs-promo artifacts moved before a clean public API model card. A BenchLM note reported Kimi 3 pages with specs, benchmarks, pricing, and weight availability still marked as coming soon.

kimmonismus framed K2.6 as the baseline K3 has to clear in a prelaunch recap: 12-13 hour autonomous coding runs, thousands of tool calls, 256K context, native vision, and up to 300 sub-agents.

The architecture rumor stack is messy enough to keep as a list:

• 1M context appeared in koltregaskes's docs-linked post and in the Kimi UI screenshot's Extra Long mode.
• 2.5T parameters came from an April Chinese tech article, but Lentils80 said Chinese users generally viewed the source as low credibility.
• eliebakouch guessed linear attention, a 6:1 ratio, DSA for global attention, 70 layers, 1.5T parameters, about 30T training tokens, latentMoE, Kimi residual attention, and native multimodality in a post that disclaimed having no information.
• teortaxesTex counter-guessed 48T tokens, 2.4T total parameters, and 40-60B active parameters in a reply, while doubting DSA-style global layers.
• eliebakouch added agent swarm, Muon optimizer, RL on Kimi Code CLI, MOPD, and low cache-hit cost in a follow-up, with a caveat about external providers caching linear-attention state.

The only hard technical surface in official docs remains K2.7/K2.6, not K3.

The reaction was less about one Arena name than about open-weight frontier pressure. daniel_mac8 argued in a policy-and-pricing post that a Fable-level Kimi K3 would be a larger DeepSeek-style shock because frontier-level intelligence would be downloadable rather than locked behind proprietary pricing.

haider1 made the weaker but more operational version in a competition post: Chinese models would not need to beat Fable or Sol if they surpassed Opus 4.8 or GPT-5.5 with strong self-hostable options.

hyhieu226 tied the expectation to the open-versus-closed gap in a commoditization post, arguing that a large K3 jump would weaken the idea that recursive self-improvement had made the gap impossible to bridge.

The commercial backdrop is not just leaderboard anxiety. deedydas compiled reported and leaked numbers across Chinese AI labs and put pure-play lab revenue run rate at $2.6B, with Moonshot listed at $200M run rate and a rumored $30B valuation.

The same chart put Zhipu at $1B run rate, DeepSeek at $500M, Kuaishou's Kling at $500M, and MiniMax at $400M. The footnote marked several figures as rumored, press-reported, or based on filings and analyst reports, which is the right level of confidence for the numbers until company disclosures catch up.