Moonshot AI releases Kimi K3

K3's headline number is 2.8 trillion parameters, but the model is built as a sparse mixture-of-experts, so only a fraction of that mass activates on any given request ^[1]. Moonshot paired that with two new mechanisms, Kimi Delta Attention (a hybrid linear-attention design) and Attention Residuals, and the combination let K3 use 21% fewer output tokens than its predecessor K2.6 on equivalent tasks ^[2]. The payoff shows up on Artificial Analysis's composite leaderboard, where K3 posted an Elo of 1,547, a 732-point jump over K2.6, trailing only Claude Fable 5 ^[3]. That is the real story under the market noise: a lab operating under three years of escalating U.S. export controls on advanced chips and lithography equipment closing most of the gap to frontier U.S. models through architecture rather than raw compute ^[2].

The scale of the reaction outran the release itself. Taiwan's benchmark index fell more than 6%, Japan's market closed down 4%, the Nasdaq had its worst session of the week at down 1.5%, and the Philadelphia Semiconductor Index is now off more than 20% from its June peak, with roughly $3.3 trillion in value lost since June 22 ^[4]. Nvidia shares slid enough to briefly hand its title of world's most valuable company to Apple ^[5]. The rout was not confined to chips: Synopsys and Cadence shares fell after reports that Kimi-based tooling designed a chip in 48 hours without proprietary EDA software, and even crypto assets slipped as the risk-off mood spread ^[6][7]. Notably, the pain was not purely a U.S.-versus-China story - domestic rivals Z.ai and MiniMax reportedly lost roughly 27% and 16% of their valuations respectively, showing investors repricing the entire open-weight model category, not just American incumbents ^[2].

K3's list price of $3 per million input tokens and $15 per million output tokens is roughly half of Anthropic's Opus 4.8 on paper ^[1]. But K3 is a verbose, reasoning-heavy model - one widely cited benchmark task, generating a simple SVG pelican, took 13,241 tokens and cost about $0.25, a figure that would look trivial on cheaper, terser models ^[2][3]. Atreides Management's Gavin Baker has put a number on the gap, estimating real-world cost can run 50-70% higher than the sticker price implies once token consumption is factored in ^[5]. Developer sentiment online echoes this directly: reactions to K3 have praised the model's raw capability while flagging that once verbosity is priced in, its effective cost per task lands close to GPT-5.6 rather than clearly undercutting it, a distinction the initial pricing comparison obscures.

The timing stings: just a week before K3 shipped, an Anthropic executive said Claude's technology was roughly six to twelve months ahead of Chinese rivals, a claim the release appeared to undercut almost immediately ^[8]. Wall Street's read is split. Bernstein's Robin Zhu called the release 'confirmatory' of a trend rather than a shock, and Morgan Stanley's Gary Yu framed it as an all-round catch-up of Chinese LLMs on size, performance, and pricing rather than a sudden leap ^[9]. Arena's CEO Anastasios Angelopoulos went further, calling it possibly the single biggest release of the year, while analyst Patrick Moorhead pushed back that the market's panic looked 'shockingly similar' to the overreaction after DeepSeek's R1 in January 2025, when Nvidia lost roughly $590 billion in a single session ^[10][11]. On social platforms, the same tension played out in real time: investor Gavin Baker framed K3 as a likely inflection point that is bad for Anthropic and OpenAI but net positive for nearly everyone else, while developer commentary pushed back that the model's real value is more modest than the benchmark charts suggest once its token appetite is accounted for. YouTube coverage of the release leaned enthusiastic, generally treating K3's benchmark parity with Claude Fable 5 and GPT-5.6 as the headline story.