Kimi K3 Is the Largest Open Model Ever, and It Is Not Cheap

Moonshot AI released Kimi K3 on July 16, 2026: a 2.8-trillion-parameter open mixture-of-experts model with a 1M-token native vision context, benchmarks that land it in the frontier conversation, and a price tag of $3.00 per million input tokens and $15.00 per million output. That last number is the story. K3 is the largest open-weight model ever announced, and it is also the most expensive model any Chinese lab has ever shipped, roughly triple Kimi K2.6's $0.95/$4. The reflex that a Chinese frontier model means a bargain just stopped being true.

• 2.8T parameters, 16 of 896 experts active per token, a 1M-token context, and native vision. Moonshot calls it the first "open 3T-class" model, taking the size crown from DeepSeek's 1.6T V4 Pro.
• Weights are not out yet. The model is live in the Kimi app and API today; Moonshot has promised the open weight release by July 27, 2026. Until then "open" is a roadmap item, not a fact.
• No SWE-bench Verified number exists. Every coding score Moonshot published comes from its own suites and its own KimiCode harness. That is why K3 enters our leaderboard unranked.
• Independent evals put it third or fourth, not first. Artificial Analysis has it third on GDPval-AA v2 behind Claude Fable 5 and GPT-5.6 Sol, and it leads Arena's Code WebDev board on blind developer preference.

What actually shipped?

K3 is a mixture-of-experts model with 2.8 trillion total parameters that activates 16 of 896 experts per token, which is how a model this large stays servable at all. Moonshot credits two internal architecture changes: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, described as a drop-in replacement for standard residual connections. Together the lab claims roughly 2.5x better overall scaling efficiency than Kimi K2. The model applies quantization-aware training from the supervised fine-tuning stage onward, using MXFP4 weights with MXFP8 activations, which is a serious engineering answer to the obvious question of how anyone runs a 2.8T model economically.

RelatedApple Tells 40 Ex-Staff at OpenAI to Preserve Data

The launch scores Moonshot published: 93.5% on GPQA Diamond, 88.3% on Terminal-Bench 2.1, 91.2% on BrowseComp, 56.0% on Humanity's Last Exam with tools, and 84.2% on MCP Atlas. On coding it reports 67.5 on DeepSWE 1.0 and 81.2 on FrontierSWE, both via its own KimiCode harness.

Why is the price the real news?

For two years the strategic logic of Chinese open-weight models was straightforward: match Western quality closely enough, give the weights away, and undercut on API price until the cost of frontier intelligence collapsed. DeepSeek built its reputation on exactly that trade. K3 abandons half of it. At $3.00 cache-miss input and $15.00 output, K3 is priced in the same neighborhood as Anthropic's Sonnet tier. Cache-hit input at $0.30 is the one place the old instinct survives.

That repricing tells you something about the economics underneath. A 2.8T model, even a sparse one activating 16 experts, is expensive to serve. Moonshot is no longer willing to eat that to win a price war, which suggests the era of Chinese labs subsidizing frontier inference to buy market share is running into the same compute bill everyone else pays. If you budgeted for open Chinese models as your cheap tier, that line item just changed.

Can you trust the benchmarks?

Carefully, and this is where we diverge from most of today's coverage. Moonshot has not published a SWE-bench Verified score for K3. The coding evals in its launch materials run on DeepSWE, FrontierSWE, Terminal-Bench and Kimi Code Bench, several of which Moonshot built. Different rows in its comparison table use different harnesses: KimiCode for K3, Claude Code or Codex for rivals. Those bars are not a controlled head-to-head, and Moonshot does not claim they are.

This is a pattern, not a one-off. As of June 2026, no independent third-party result existed for Kimi K2.7 on SWE-bench Verified, SWE-bench Pro, or Terminal-Bench 2.0 either. Every number came from the lab. So on our AI Coding Leaderboard, K3 goes in unranked with a "verifying" chip and no printed score, alongside GPT-5.6 Sol and Grok 4.5, until an independent evaluation lands. Kimi K2.6 keeps its rank at 80.2% because that number is checked. We would rather show a blank than a number we cannot stand behind.

The independent signals that do exist are genuinely good, and worth separating from the vendor table. Artificial Analysis scored K3 at 1,687 on GDPval-AA v2 across 44 occupations, third overall 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. On AA-Briefcase, a private agentic long-horizon benchmark, K3 placed second at 1,527, beating GPT-5.6 Sol Max at 1,495. In Arena's blind testing, developers preferred Kimi over every leading US model for front-end coding. Third or fourth on the frontier, from an open lab, is a real result. It is just not "beats everything," which is how the vendor table reads at a glance.

What does it mean for the market?

The signal for investors is that the cheap-open-Chinese-model thesis, which has been used to argue frontier inference margins are doomed, just got complicated by the lab that should be proving it. Moonshot pricing a flagship near Anthropic's Sonnet tier is evidence that serving 3T-class models costs real money regardless of where the lab is domiciled, which is directionally supportive of the pricing power at OpenAI and Anthropic that the bear case says evaporates.

The clearer exposure is compute. A Chinese lab shipping a 2.8T model trained and served under export controls is a demand story for accelerators and memory, not a substitution story, and it lands the same week TSMC raised 2026 capex to $60-64B and ASML raised its outlook. Watch whether K3's weights actually drop on July 27: a genuinely open 2.8T checkpoint is a step-change in the amount of GPU and high-bandwidth memory that self-hosters and neoclouds need to run frontier-class inference. This is analysis, not investment advice.

Related29 Countries Sign China-Led AI Governance Body in Shanghai

Who should care today?

If you build on open weights, the honest answer is wait until July 27. Right now K3 is an API product with an open-weights promise attached, and a 2.8T checkpoint is not something most teams can self-host anyway; realistically it lands as a neocloud offering, not something on your own hardware. If you are picking a coding model, K2.6 at $0.95/$4 remains the better-understood value and still holds a verified 80.2% SWE-bench Verified. If you run long-horizon agentic work, the AA-Briefcase second-place finish is the most interesting number in this launch, because it is independent, private, and measures the thing agent builders actually care about.

One practical warning from Simon Willison's launch-day testing: K3 currently ships with a single reasoning effort setting, "max," and it shows. His pelican SVG test burned 13,241 reasoning tokens to produce 3,417 output tokens, costing 25 cents for one image. At $15 per million output, a model that cannot dial down its own thinking is a budget hazard on trivial calls.

1. Jun 15, 2026No independent SWE-bench result exists for Kimi K2.7 every published score is vendor-run
2. Jul 16, 2026Kimi K3 announced: 2.8T params, 1M context, $3/$15 API and app only
3. Jul 17, 2026Artificial Analysis places K3 third on GDPval-AA v2 first independent read
4. Jul 27, 2026Promised open weight release the claim that makes "open" true

Our take

K3 is the most impressive open-weight launch of the year and the least generous one, and those two facts are the same fact. Moonshot built something genuinely frontier-adjacent, third on the most credible independent eval available, first among open labs, and then priced it like a company that has read its own compute invoice. We think that is the honest move, and we think it quietly retires the narrative that Chinese labs will drive the cost of intelligence to zero. They will drive it down until the physics bill arrives.

The part we will not go along with is the benchmark table. A launch that leads with scores from harnesses the vendor wrote, against rivals run on different harnesses, is a marketing artifact, and the absence of any SWE-bench Verified number in a launch this loud is conspicuous. K3 does not need it: the Artificial Analysis and Arena results are strong and independent. Lead with those. Until an outside lab runs the coding suites, our leaderboard will keep K3's score column blank, and we would encourage you to treat the vendor bars the same way.

Original analysis by GenZTech. Reporting on the K3 launch via VentureBeat and Axios.