Yesterday, July 16, Moonshot AI announced Kimi K3: a 2.8 trillion parameter mixture-of-experts model with a one million token context window, native image and video input, and a commitment to publish the weights on July 27. It is the largest open-weights model ever announced, and independent scoring already places it fourth overall among all current models. Here is the plain-English version, including the parts the launch coverage glosses over.
The facts
- Released: July 16, 2026, API-only for now. Weights and the license text are promised for July 27.
- Size and design: 2.8T total parameters, with 16 of 896 experts active per token. Only a fraction of the model runs on each request, which is how something this large stays affordable.
- Context: about one million tokens. Entire codebases, contract sets, or years of correspondence in one request.
- Price: $3 per million input tokens and $15 per million output on Moonshot's API, with cached input dropping to $0.30. That undercuts Western flagships on sticker price.
- Standing: Artificial Analysis, an independent evaluator, scores it 57 on its Intelligence Index, fourth of the 189 models it tracks.
The honest benchmark picture
The launch charts look spectacular. Read them with these three adjustments:
- Vendor and independent numbers disagree. Moonshot reports 88.3 on Terminal-Bench, a demanding agentic coding test. An independent re-run measured 80.9. Both are strong; they are not the same claim.
- The classic yardsticks are missing. Moonshot did not publish SWE-bench Verified or SWE-bench Pro results, the scores most 2026 coding comparisons are built on, and substituted newer benchmarks instead. That is increasingly common across the industry and it makes direct comparisons harder, which is sometimes the point.
- It is verbose. During independent evaluation K3 produced roughly twice the median output tokens. You pay per token. A model with cheap-looking rates and chatty answers can cost more per task than a pricier, tighter one.
One result that is independent and genuinely striking: K3 currently ranks first on LMArena's frontend coding leaderboard, ahead of every Western flagship. For web interface work, it has real claims.
Why open weights matter to a business that will never self-host
Most of our clients will never run a 2.8 trillion parameter model on their own hardware. The weights still matter to them, for three reasons.
First, price gravity. When frontier-class weights are public, hosting providers compete to serve them, and API prices across the whole market feel the pull. Second, continuity. A model whose weights you can obtain cannot be taken away from you by a pricing change or a product cancellation. Third, privacy options. Open weights are what let providers like Ollama host large models with no-logging commitments, and what make fully private deployments possible for the workloads that genuinely need them.
The license is the fine print that decides everything. "Open" is a spectrum. Moonshot's earlier K2-family releases carried usage clauses that required commercial deals at large scale, and K3's actual license text does not exist publicly until the weights ship on July 27. If you are planning to build on K3 commercially, wait and read it. We will.
What to actually do
For most businesses: nothing yet. A model that is four days old, API-only, and awaiting its license is a thing to watch, not a thing to migrate to. If you run meaningful API volume, put K3 on your evaluation list for August, after the weights and license land and independent numbers settle. Test it on your own workload, not on benchmarks: your invoice-matching prompt is the only benchmark that pays rent.
We have already added K3 to our LLM Ranking Tool with independently verified figures where they exist and estimates clearly flagged where they do not. As of today it ranks fifth for coding-assistant workloads on our default weights. If you want help deciding whether any of this belongs in your stack, that is a conversation we have every week.