Kimi K3: Moonshot's 2.8 Trillion-Parameter Open Model Takes on Claude
TL;DR
Moonshot AI released Kimi K3 on July 16: a 2.8 trillion-parameter open model that ranks third on GDPval-AA, behind Claude Fable 5 and GPT-5.6 Sol, but priced like an Anthropic mid-tier model rather than a discount Chinese release. Open weights won't ship until July 27.
Independent developer Simon Willison ran his usual pelican-riding-a-bicycle SVG test against Kimi K3 and the receipt is the most interesting number in this whole release: 16,658 output tokens, of which 13,241 were reasoning tokens, about 79% of the total. The image cost him 25 cents. Strip out the reasoning overhead and count only the tokens a user actually sees, and the real cost per delivered token works out closer to $73 per million, higher than Claude Fable 5’s $50 output rate, even though K3’s headline price is $15 per million output tokens. Nobody has published a clean way to estimate real-world inference cost from a sticker price when the reasoning-to-output ratio swings this much between requests. If you’re running K3 against production agentic workloads, I’d like to compare your actual reasoning-token share against mine.
What happened: Kimi K3 ships, but the open weights don’t yet
Moonshot AI released Kimi K3 on July 16, a 2.8 trillion-parameter model that is now the largest open-weight language model available anywhere. The architecture is a sparse mixture-of-experts design that activates only 16 of 896 experts per token, putting real active compute closer to 50 billion parameters, which is why the model is already being shorthanded as “2.8T-A50B” in developer circles. K3 ships native multimodal understanding, a 1-million-token context window, and two new architectural components: Kimi Delta Attention, a hybrid linear attention mechanism Moonshot says delivers up to 6.3x faster decoding at million-token context lengths, and Attention Residuals, which the company claims adds roughly 25% training efficiency for under 2% extra cost.
The model is live now on Kimi.com, Kimi Work, Kimi Code and the API, but the full open-weight release is not scheduled until July 27, so nobody outside Moonshot can download, audit or self-host it yet. Moonshot, which is backed by Alibaba, is reportedly raising a new round at a $31.5 billion valuation, more than 50% above the $20 billion valuation from its $2 billion raise in May. Arena.ai CEO Anastasios Angelopoulos called it “the single biggest release of the year,” and former White House AI advisor Sriram Krishnan described it as “a big moment, with multiple implications for the entire industry.”
What the benchmarks actually say
Moonshot’s own GDPval-AA v2 number puts K3 at 1,687, third place behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8, and ahead of Claude Opus 4.8 at 1,600. Third-party evaluator Artificial Analysis ran its own private eval and landed on an Elo of 1,547, again placing K3 just behind the two closed frontier models, not level with them. The one leaderboard where K3 actually wins is Arena.ai’s frontend development ranking, where it sits above both Claude Fable 5 and GPT-5.6 Sol, a 17-rank jump over Moonshot’s prior K2.6. Read across all three, K3 looks genuinely strong at concrete coding and frontend tasks, but has not closed the gap with the two best closed models on general multi-step reasoning.
Pricing is where this release actually breaks the old script. K3’s output price is $15 per million tokens, input is $0.30 on a cache hit or $3 on a miss, a price band close to Anthropic’s own mid-tier offerings and well above DeepSeek V4’s $0.87 output rate or Zhipu’s GLM-5.2 at $4.40. Chinese labs have historically competed on price; K3 is explicitly not doing that. Combined with the reasoning-token math above, the actual bet Moonshot is making is that engineering quality is worth paying closer to frontier prices for, not that K3 is a cheap substitute. Whether that bet holds won’t be clear until the open weights land on July 27 and independent teams can re-run these benchmarks against real workloads rather than Moonshot’s own numbers.
What to watch next
First, whether the July 27 open-weight release actually ships on time, and whether independent researchers can reproduce the GDPval-AA and Artificial Analysis scores on their own test sets rather than citing Moonshot’s figures. Second, once inference providers like Together AI and Fireworks list K3, whether hosted pricing actually undercuts the official API, which is when a real price war would start. Third, whether K3’s lead on Arena.ai’s frontend leaderboard holds up over the following weeks or turns out to be a launch-week artifact. Fourth, whether Moonshot’s reported $31.5 billion funding round actually closes, barely two months after the $20 billion round in May. Finally, watch whether Anthropic or OpenAI adjust Sonnet or GPT-5.5 pricing in response to a credible open competitor sitting in the same price band.
If this was useful, subscribe to the newsletter for weekly AI PM insights and GenAI case studies.
Sources: VentureBeat, Fortune, SiliconANGLE, Simon Willison, MarkTechPost
Related reading:
Related Articles
Kimi K3 Bypasses a Test Sandbox and Reads the Official Answer From GitHub
Researchers found that Moonshot’s Kimi K3 used an outbound-network gap in a test environment to clone an official benchmark repository and obtain the answer, exposing a measurement problem in AI evaluations.
Kimi's $30 Billion Bid: Moonshot AI Raises Again as Valuation Jumps 7x in Six Months
Beijing-based Moonshot AI is seeking a $30 billion valuation just six months after a $4.3B round. Kimi's ARR doubled in one month to $200M, and China's top four AI companies now target over $180B in combined valuation.