Moonshot's Kimi K3 Sets Open-Weight Epoch Record and Tops 3D Design Benchmark
Moonshot's Kimi K3 model set a new open-weight record on the Epoch Capabilities Index and ranked first in the DesignArena 3D design benchmark with an Elo of 1450. The model also placed second on the Artificial Analysis AA-Briefcase agentic work benchmark with an Elo of 1543, trailing only Claude Fable 5's 1574 and outperforming GPT-5.6 Sol, Claude Sonnet 5, and Claude Opus 4.8.
Performance gains stem from a new transformer architecture that replaces fixed residual connections with attention-based mixing across network depth to stabilize training at scale. Benchmark performance requires high computational effort, with each AA-Briefcase task averaging 56 minutes and $10.57 to run, roughly 10 times the cost of the previous generation due to longer reasoning cycles and heavier output token generation. The model's API is already live, with weights scheduled for public release on July 27.
From the sources (25 posts)
@designarenaBREAKING: Kimi K3 by @Kimi_Moonshot is officially 1st on Frontend Web App Arena by DesignArena With an Elo of 1326, this open-weight model leads the way, ahead of Fable 5, Sonnet 5, and Opus 4.8 by @AnthropicAI Huge congrats to the @Kimi
@kimi_moonshotRT @DesignArena: BREAKING: Kimi K3 by @Kimi_Moonshot is officially 1st on Frontend Web App Arena by DesignArena With an Elo of 1326, this…
@crystalsssupKimi is #1 in Design Arena, a benchmark for frontend website-building capabilities.
@arenaExciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 and GPT-5.6 Sol. If Kimi K3's weights are released on schedule by July 27, it will become the #1 open-weight model. This release marks a maj
@arenaKimi K3 ranks #4 overall (+9.6%) - #1 Confirmed Task Success (+14.4%) - #3 Praise vs. Complaint (+20.6%) - #4 Tool Hallucination (+1.1%) - #14 Steerability (+5.6%) - #17 Bash Recovery (+6.4%)
@scaling01RT @arena: Exciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 and GPT-5.6 Sol. If Kimi K3's…
@kimi_moonshotRT @arena: Exciting update: Kimi K3 has landed at #4 on the Agent Arena leaderboard, matching Claude Opus 4.8 and GPT-5.6 Sol. If Kimi K3's…
@rohanpaul_aiKimi K3 is ahead of Claude Feble 5 again. Has taken the top spot on DesignArena's Frontend Web App benchmark. On DesignArena bench, AI models receive the same app-building prompt, produce competing interfaces, and users vote for the bette
@epochairesearchMoonshot's Kimi K3 scores 156 on the Epoch Capabilities Index (ECI), setting a new open-weights record. This places it between Opus 4.6, and GPT 5.4, which released in February and March 2026 respectively, and just ahead of GPT 5.6 Luna. ht
@scaling01RT @EpochAIResearch: Moonshot's Kimi K3 scores 156 on the Epoch Capabilities Index (ECI), setting a new open-weights record. This places it…
@teortaxestexKimi K3 scores 5'11 3/4" on Epoch Capabilities index, ahead of GPT 5.6 Luna at 5.11" GPT 5.6 Terra (Gigachad) reigns supreme at 6' (actually 155.6 vs 155.3 vs 158.4) @scaling01 takes the W
@zacharynadoRT @peterwildeford: Kimi K3 is exactly the level of capability you would predict it to have given the long-term 2 year trend of Chinese AI…
@kimi_moonshotRT @DesignArena: BREAKING: Kimi K3 by @Kimi_Moonshot is 1st overall on 3D Design with an Elo of 1450. This is a 6 position and 108 Elo jum…
@crystalsssupKimi K3 is #1 in 3D design arena
@teortaxestexThis is absurdly impressive. BrokenArXiv is *hard* for LLMs and OpenAI focused on such problems hard and holds a commanding lead. Kimi is not a frontend slop machine, it's a generalist proto-AGI (though same can be said of Meta, congrats) h
@scaling01my prediction for GPT-6's ECI is 166.5 (given that it's released at the end of august) so roughly ~10 ECI points higher than Kimi-K3 depending on which trendline you take this implies US frontier labs are ahead 8.2 - 10.9 months
@teortaxestexI don't much like routing but it seems it's starting to work
@artificialanlysKimi K3 is second only to Fable 5 on AA-Briefcase, our agentic knowledge work benchmark, but costs more than Opus 4.8 to run while averaging nearly an hour per task Last week @Kimi_Moonshot released Kimi K3, a 2.8T parameter model that sco
@artificialanlysAA-Briefcase measures model performance across three dimensions: binary rubric checks for ground-truth correctness, pairwise grading on analytical quality, and pairwise grading on presentation quality. The AA-Briefcase Elo is a single metri
@artificialanlysKimi K3’s frontier performance comes at a high Cost per Task. Its average Cost per Task of $10.57 is one of the highest recorded, below Claude Sonnet 5 (max, $14.43) and Claude Fable 5 ($22.30). Compared to Kimi K3, GPT-5.6 Sol (max) trails
@artificialanlysKimi K3 has an average Time per AA-Briefcase Task of 56.4 minutes. This is driven by a high number of turns, as well as higher output token use and lower speeds using the first party Kimi API. Kimi K3 uses 120k output tokens per task and 83
@kimi_moonshotRT @ArtificialAnlys: Kimi K3 is second only to Fable 5 on AA-Briefcase, our agentic knowledge work benchmark, but costs more than Opus 4.8…
@rohanpaul_aiMoonshot's Kimi K3 became strongest open-weights model ever measured on Epoch Capabilities Index (ECI). ECI blends dozens of separate benchmarks into a single score, so one benchmaxxed test result can't distort the ranking the way it does
@ylecunRT @JitendraMalikCV: Highly performant open weights frontier models such as Kimi are a competitive threat to OpenAI & Anthropic, but probab…
@rohanpaul_aiKimi K3 just grabbed another crown. 1st overall on 3D Design with an Elo of 1450.