Moonshot AI, the Beijing lab behind the Kimi model family, released Kimi K3 on July 16, 2026 — 2.8 trillion parameters, a 1 million-token context window, and native vision. It is the largest open model ever released, 75% bigger than DeepSeek V4-Pro.
Within hours it took #1 on LMArena’s Frontend Code Arena at 1,679 points, a 17-place jump from K2.6, and beat Claude Fable 5 in 76% of head-to-head frontend matchups.
The full weights arrive July 27 under an MIT license. That combination — frontier-adjacent quality, fully open weights, permissive license — is what makes kimi k3 moonshot 2.8 trillion the most consequential model release of the quarter, regardless of where it eventually settles on the leaderboards.
Where Kimi K3 actually ranks
K3 is #1 in frontend code and #4 in general intelligence. Both numbers matter, and they are not the same claim.
| Benchmark | Result |
|---|---|
| LMArena Frontend Code Arena | #1 — 1,679 points (up 17 places from K2.6) |
| Frontend domain breakdown | #1 in 6 of 7 domains |
| Head-to-head vs Claude Fable 5 (frontend) | Wins 76% of matchups |
| Artificial Analysis Intelligence Index | #4 globally at 57.1 |
| Ahead of K3 on that index | Claude Fable 5 (59.9), GPT-5.6 Sol (58.9) |
The 2.8-point gap to Fable 5 on Artificial Analysis is real but narrow. Bloomberg Intelligence put the US-China model gap at 6% in June 2026; K3 closes it further. MegaOne AI’s earlier analysis found Fable 5 charging roughly twice the price for 5.7% more capability — K3 attacks precisely that price-per-point argument.
The architecture: 1.8% of experts active per token
K3’s efficiency comes from three named components rather than raw scale.
- Kimi Delta Attention — the successor to the attention scheme in K2, tuned for the 1M-token context.
- Attention Residuals — residual pathways across attention layers that stabilize training at trillion-plus scale.
- Stable LatentMoE — a mixture-of-experts router activating just 16 of 896 experts per token, about 1.8%.
Moonshot reports 2.5x scaling efficiency versus K2 from this combination. The sparsity figure is the headline: a 2.8T-parameter model that activates under 2% of its weights per token has an inference cost closer to a mid-size dense model than its parameter count suggests.
Training ran on export-grade Nvidia silicon plus an unnamed alternative GPU vendor. The hardware provenance is contested — the White House has claimed Moonshot used banned Nvidia chips, a dispute MegaOne AI covered in its report on the export-control allegation.
Demand broke the signup queue
Moonshot halted new signups on July 20, four days after launch, because demand exceeded serving capacity.
One documented case study had K3 running autonomously for 48 hours designing a simulated system — the kind of long-horizon agentic run that consumes enormous token volume per task and explains why capacity ran out so fast.
The signup halt is a supply constraint, not a demand signal to discount. It also makes the July 27 weights release more significant: anyone blocked from Moonshot’s API can self-host instead.
What it did to Chinese AI stocks
The market read K3 as a consolidation event, not a sector win.
- Zhipu fell 28.49%
- MiniMax fell 15.62%
Both are direct competitors in China’s open-weight tier — MiniMax released M3 with a 1M-token context earlier this year, which MegaOne AI covered when it rivaled top labs on context length. K3 matched the context and beat them on quality.
One analysis pegged the combined valuation impact on OpenAI and Anthropic at $314 billion. Treat that figure as a modeled estimate on private marks rather than a realized loss — but the direction is defensible, because it prices the open-weight discount rather than a capability deficit.
The economics: -5 per million tokens versus -60
Self-hosted open weights run roughly $1-5 per million tokens at scale. Proprietary frontier APIs run $15-60 per million. That is a 5x to 15x spread for a model 2.8 points behind the leader on a general-intelligence index and ahead of it on frontend code.
The spread only converts to savings above a volume threshold. Serving a 2.8T-parameter MoE requires substantial GPU memory even at 1.8% activation, so self-hosting has a fixed cost floor that kills the argument for low-volume workloads. The decision rule:
- High-volume, code-heavy, latency-tolerant — K3 self-hosted wins on cost, and now on frontend quality too.
- Low-volume or spiky — proprietary APIs still win; you are renting the GPU amortization.
- Regulated or air-gapped — open weights win regardless of price, because the alternative does not exist.
The third case grew sharply this quarter. Anthropic blocked Chinese firms from Claude Code, and the Fable 5 export episode showed that a proprietary model can be switched off by government order in hours. Weights on your own hardware cannot be revoked by letter.
What to do before July 27
If you are evaluating K3, benchmark it on your own frontend workload rather than trusting the arena score — LMArena measures human preference on generated interfaces, which correlates with but does not equal correctness on your codebase.
If you plan to self-host, size the deployment now. A 2.8T MoE at 16/896 activation still needs the full parameter set resident or sharded across nodes; the activation ratio cuts compute, not memory. Model cards and weights will appear on Moonshot’s Hugging Face organization, and live rankings are at LMArena and Artificial Analysis.
MegaOne AI tracks 139+ AI tools across 17 categories, and the open-weight tier is the fastest-moving segment in that database.
