Kimi K3 Ships With 2.8 Trillion Open Weights. The Open Frontier Ceiling Just Went Up 8x in Three Days.
Moonshot AI put Kimi K3 live on Thursday, July 16, 2026. It is a 2.8 trillion parameter Mixture-of-Experts model with a 1 million token context window, native vision, hosted API pricing at $3.00 input and $15.00 output per million tokens (with a $0.30 cache-hit rate), and full weights promised under a Modified MIT license by July 27. Two variants went live: K3 Max for chat and agent work, K3 Swarm Max for large-scale parallel processing. The weights are not on Hugging Face yet, but Moonshot has run this play before and shipped K2 on the same license, so the ten-day countdown is the real news.
Three days ago the open ceiling was Z.ai's GLM-5.2 at roughly 355 billion total parameters. Kimi K3 is roughly 8x larger by total params, roughly 8x by context length, and it clears GLM on almost every coding benchmark Moonshot chose to publish. That is the story: the open frontier ceiling just moved, again, in three days, and this move is the biggest one open weights have ever made.
The Numbers
Every benchmark below is vendor-reported. Neutral-harness replication has not landed on Kimi K3 yet and everything Moonshot showed at launch should be treated the way we treat any first-day vendor scoreboard, which is skeptically. That said, here is the shape of what the vendor is claiming.
| Field | Kimi K3 | Notes |
|---|---|---|
| Total parameters | 2.8T | Sparse MoE, 896 experts, 16 routed per token |
| Active per token | ~50B | 16 of 896 experts, roughly 1.8% activation |
| Attention | KDA | Kimi Delta Attention, hybrid linear, plus Attention Residuals |
| Context window | 1M tokens | Native, up from 128k on GLM-5.2 |
| Vision | Native | Not a bolt-on adapter |
| Hosted price | $3 / $15 | Per million tokens, input / output; $0.30 cache-hit input |
| Weights | Modified MIT | Public download promised by July 27 |
| DeepSWE (vendor) | 67.5 | Beats Opus 4.8 and GPT-5.5 on Moonshot's harness |
| Terminal-Bench 2.1 | 88.3 | Vendor claim, leads the frontier board |
| FrontierSWE (vendor) | 81.2 | Trails Fable 5, in range of Opus 4.8 per Moonshot's numbers |
Source: Moonshot AI launch blog, Kimi API platform docs, third-party model cards. Neutral-harness replication pending.
Two things worth pinning down. First, active parameters. Sixteen of 896 experts fire per token, so active weight count lands near 50 billion, not 2.8 trillion. GLM-5.2 routes about 32 billion active. Active-to-active, Kimi K3 is closer to 1.6x GLM, not 8x, and that is the number that maps to per-token inference cost. Second, capability is not a linear function of parameters. If Moonshot's benchmarks survive independent harnesses, and that is the only sentence in this piece that is still an open question, Kimi K3 is genuinely in Opus and Fable territory on coding. If they do not survive, Kimi K3 is still the largest open MoE ever released, and the numbers below still hold.
The Three-Day Ramp
On July 13, we wrote up GLM-5.2 as the sharpest open-weights release of the year: fourth overall on the Artificial Analysis Intelligence Index, priced 82 percent below Opus 4.8, and reportedly moving 40 percent of developer tokens on OpenRouter. That piece opened by noting the ceiling had moved. Three days later the ceiling moved again, by a factor of eight, from a different Chinese lab, with a technical report that keeps getting sharper on architecture efficiency.
| Ship date | Model | Total | Active | Context | Hosted price |
|---|---|---|---|---|---|
| Feb 2026 | DeepSeek V4 | ~250B | ~25B | 128k | $0.14 / $0.28 |
| Jul 13 | GLM-5.2 | ~355B | ~32B | 128k | $1.40 / $4.40 |
| Jul 16 | Kimi K3 | 2.8T | ~50B | 1M | $3.00 / $15.00 |
Source: vendor announcements. DeepSeek V4 pricing at DeepSeek platform rates.
The trajectory is not subtle. Total params climbed 11x in five months. Context length climbed 8x in three days. Hosted output price climbed too, but even at $15, Kimi K3 is 70 percent below Opus 4.8 on output and 40 percent below GPT-5.6 Sol. The open frontier is now dictating both a capability ceiling and a price ceiling to the closed frontier. That is a different market than the one open weights lived in six months ago.
The Self-Host Math
Open weights are only sovereign if you actually run them. On GLM-5.2, we did this math and it came out at roughly 1.5 TB of GPU memory at full precision, or nineteen H100s. Kimi K3 makes that number look manageable. Here is what full-precision hosting looks like at 2.8 trillion parameters.
| Precision | Weight memory | H100 80GB units (weights only) | MI300X 192GB units (weights only) |
|---|---|---|---|
| fp16 / bf16 | ~5.6 TB | 70 | 30 |
| int8 | ~2.8 TB | 35 | 15 |
| int4 | ~1.4 TB | 18 | 8 |
Weight-memory only. Add roughly 30 to 50 percent for KV cache, activations, and overhead at production batch sizes. 1M-token context balloons KV cache further and usually forces a second cluster tier just for long-context serving.
Read the fp16 line. Full-precision Kimi K3 needs a cluster north of seventy H100 80GB cards just for weights, before you cache a single token of context. At list price that is roughly a $2 million capital budget and a data center power slot, and once you add the KV cache pressure of a 1M-token window at real batch sizes, the operational number gets worse fast. Even int4, which trades capability for footprint, still needs about eighteen H100s. There is not a serious universe where a startup or a small enterprise hosts this locally. Anyone who says otherwise is quoting the weights-only column and hoping you skipped the footnote.
The Sovereignty Problem, Sharper
Three days ago the argument was that GLM-5.2 was legally open and practically hosted. Kimi K3 makes that argument tighter, not weaker. When the model gets bigger, the fraction of users who route through the vendor's own cloud gets closer to one, because the fraction of teams who can afford to host it locally trends the other way. Kimi's own API endpoint is the path of least resistance for almost everyone using K3 in production on day one, and Kimi's API endpoint is subject to China's National Intelligence Law, which requires Chinese companies to hand over data on government request. That fact has not moved. What moved is the size of the fraction it applies to.
The three options are the same three we laid out on GLM. Self-host at full precision and eat the capital and power bill. Route through a Western reseller (Fireworks, Together, DeepInfra, or the OpenRouter passthrough) once one of them stands up a K3 endpoint, and get your prompts processed on infrastructure you do not control but at least outside Chinese jurisdiction. Route through Kimi's own API and accept the jurisdiction cost because the price is right. On K3 the third option is going to be roughly 80 to 90 percent of real traffic, and the enterprise buyers doing the routing will call this "open source" in their board decks. The frontier labs know this, and it is why the trust moat is still worth more per token than the capability moat, even when the capability delta narrows.
What This Does to the Closed Premium Tier
Two things, running on different clocks.
On the price clock, the compression continues. Kimi K3 at $3 input and $15 output is not the cheap tier. It sits above the commodity floor we watched land last week (GPT-5.6 Luna at $1 input, Grok 4.5 at $2 input) and well above the DeepSeek and GLM lines underneath. But it is more than 5x cheaper than Opus 4.8 on output. Anthropic and OpenAI both hold their premium tiers by pointing at capability, and now the capability argument at $3 input has a Chinese open-weights answer with a technical report and a weights drop scheduled for ten days from now. The public price sheet does not have to move. The negotiating floor already did.
On the capability clock, the picture is more nuanced. Vendor benchmarks say Kimi K3 trails Fable 5 on FrontierSWE and DeepSWE and beats it on Terminal-Bench and SWE Marathon. Every one of those numbers was picked by Moonshot to be shown. Fable 5 replicated on neutral harnesses at a premium ceiling above every open number in the market; our scoreboard piece from a week ago still has Mythos 5 and Fable 5 leading SWE-Bench Pro by fifteen points. If Moonshot's numbers hold on independent replication, the premium ceiling gets shaved by another five points. If they do not, the leaderboard shifts back to a familiar pattern where the premium closed models keep the top slot and open weights own the price-per-capability curve. Either outcome pressures the closed premium tier. The first outcome is louder.
Our Take
The interesting thing about the last three days is not that Moonshot shipped a big model, because someone was going to. It is the cadence. GLM-5.2 on July 13, Kimi K3 on July 16. Two labs, three days, and the open ceiling went from 355B to 2.8T. That is not one team racing the frontier. That is a supply chain running at cruising speed, and it is running on the Chinese side of an export-control wall the U.S. spent two years building. The buyer-side outcome, if you are a builder, is that the price ceiling on frontier-class intelligence is falling faster than our 2026 pricing war piece projected, and the capability floor at $3 input has just been reset to something uncomfortably close to what Anthropic and OpenAI charge premium prices for.
The seller-side outcome is that the two open Chinese labs and DeepSeek before them have collectively shortened the useful life of a closed capability lead from months to weeks. The premium closed tier still exists, still gets paid, still runs the enterprise book. But the gap window inside which a closed lab can charge Opus prices for Opus capability is now visibly shorter than the gap window last quarter, and every open ship date compresses it further.
Three signposts I am watching from here. First, whether the July 27 weights actually land on time, in full precision, and on Hugging Face. K2 did. K3 probably will. If it slips, we downgrade the sovereignty math accordingly. Second, whether a neutral harness (Aider, TerminalBench's public leaderboard, LMArena's coding split) confirms the vendor numbers or shaves them by five to ten points, which is the usual haircut. Third, whether Anthropic or OpenAI answer with a price move at the premium tier, or hold the line and lean into the trust moat. The first tells you they see the pressure; the second tells you they see the pressure and think they can outrun it.
Every closed frontier lab has a plan for beating a closed competitor on capability. Not every closed frontier lab has a plan for beating an open competitor on price. That is the fight the next quarter is going to be about.