Kimi K3 Review: What Is Real vs Vendor-Reported (Launch-Day Test)

Moonshot AI announced Kimi K3 on July 16, 2026 as the largest open-weight model unveiled to date — roughly 2.8 trillion parameters, a sparse Mixture-of-Experts design, a 1M-token context window, and a headline table showing it beating Claude Opus 4.8 and GPT-5.5. Most write-ups are republishing that same vendor table plus one arena result. We did something different: we ran K3 against our hidden-test coding harness on launch day. This review separates what is independently verified from what is vendor-reported, then adds our own first-party numbers so you can decide with clear eyes. As of July 17, 2026.

What Kimi K3 actually is

Kimi K3 is Moonshot AI's July 16, 2026 flagship, and the pitch is size. Multiple reputable outlets (Fortune, MarkTechPost, Tom's Hardware, CNBC, Simon Willison) describe it as the largest open-weight model announced so far, at roughly 2.8 trillion total parameters. Fortune printed 2.7T; 2.8T is the figure across most primary and secondary coverage, so treat it as approximately 2.8 trillion and do not over-specify.

Architecturally it is a sparse Mixture-of-Experts model that activates 16 of 896 routed experts per token, so only a fraction of those parameters fire on any given step. The attention stack is the interesting part: Moonshot calls it Kimi Delta Attention (KDA), a hybrid linear-attention scheme — linear attention for most layers with periodic full-attention layers, plus Attention Residuals — reported to deliver up to about 6.3x faster decoding at 1M context. The window is 1M tokens, and the model accepts multimodal input (text, image, video). If you have followed Moonshot's trajectory, this is the successor lineage to the models we covered in our Kimi K2.7-Code review and Kimi K2 Thinking write-up.

The headline claim, repeated everywhere, is that on Moonshot's own tests K3 beats Claude Opus 4.8 and GPT-5.5 on most benchmarks and lands just under the top two, Claude Fable 5 and GPT-5.6 Sol. That is the sentence you should slow down on, because it is vendor-reported — and this review is mostly about telling vendor claims apart from verified ones.

Vendor vs independent vs first-party

Almost every launch-day article collapses three very different kinds of evidence into one number. We keep them in separate columns. That separation is the single most useful thing you can carry out of this review.

The correction worth stating plainly: it is not true that no independent benchmarks exist. Two aggregate independent signals do — the Frontend Code Arena and the Artificial Analysis Index. What is still missing is an independent, standalone coding suite (SWE-bench Verified, LiveCodeBench) run by someone other than Moonshot. Artificial Analysis, notably, also calls K3 more expensive and more verbose than its price-tier peers, which lines up with what we saw. For the wider field, see our best open-source LLM of 2026 guide and the overall best LLM 2026 ranking.

Our launch-day executed run

Here is the part nobody else has. On K3's launch day we ran it through OpenRouter using our nine-task, generate-code-then-run-hidden-tests harness — the same one behind our 13-model coding cost benchmark.

Result: K3 passed every task that actually executed — 8 distinct tasks across two runs (two_sum, valid_parentheses, merge_intervals, roman_to_int, lcs_len, flatten, top_k_words, token_bucket), all correct. The ninth task, parse_csv_line, never completed: the day-old model was repeatedly rate-limited upstream (HTTP 429) across all retries in both runs. To be scrupulous: that is a launch-day capacity signal, not a capability miss. We are not claiming 9/9, and we are not calling the ninth a clean failure — it simply could not be evaluated.

The calibration that matters: in our 13-model benchmark, 10 of 13 models already score a perfect 9/9 on this suite. So a frontier model clearing every executable standard task is table stakes, not a headline. Do not let anyone sell you single-shot correctness on common tasks as a differentiator. The real, defensible signals from launch day are three:

• Cost is frontier-tier. At $3-in/$15-out, one run computed $4.37 per 1,000 tasks — essentially Claude Opus 4.8 territory ($4.05/1k in our benchmark), and roughly a 3x jump from K2.6's much cheaper generation.
• It is token-lean for a reasoner. About 150 reasoning tokens per task, leaner than Kimi K2.7-Code's 272 — though note Simon Willison's launch-day pelican prompt burned ~13,241 reasoning tokens, so verbosity is prompt-dependent, and Artificial Analysis flagged it as verbose overall.
• The independent Arena win is the story. Blind human preference, not a self-report.

Where K3 lands on cost

Because correctness on standard tasks is a wash across frontier models, cost is where the decision actually gets made. The chart below places our measured K3 cost against Claude Opus 4.8 and the broad field of high-scoring models from our benchmark.

The takeaway is uncomfortable for a hype cycle: K3 is excellent and expensive. If your workload is standard coding, you are paying frontier rates for output that cheaper models also nail. K3 earns its price on the hard tail — long-horizon agentic work and frontend generation, where the independent Arena result actually differentiates it.

Pricing, access and open weights

API pricing is $3 per 1M input, $15 per 1M output, and about $0.30 per 1M cache-hit input, flat across the full 1M context. On OpenRouter the slug is moonshotai/kimi-k3, currently served by a single provider, and reasoning_effort only supports max right now. For the full Moonshot price history and how K3 compares to earlier Kimi releases, see our Kimi API pricing guide. K3 does not have a dedicated model page on our gateway yet, so pricing and access live on our pricing page.

On open weights, be careful. Some promotional coverage phrases K3 as the largest open-weight model available today. It is not available today. The weights are a promise for July 27, 2026 under a Modified MIT license, and as of July 17 they are not downloadable. So “open” is a near-term commitment, not a fact you can build on this week. If self-hosting is your reason to care, wait for the drop and verify the license terms then.

Should you use it (a decision rule)

Here is the checklist we would actually apply:

• Everyday coding, cost-sensitive? Skip K3 for now. A 9/9 model at a fraction of $4/1k does the same job. Use our cost benchmark to pick one.
• Hard frontend or long-horizon agentic work? Try K3. The blind Frontend Code Arena win is the one signal here that is both independent and directly relevant.
• Need reproducible per-suite proof before you commit? Wait. Independent standalone SWE-bench Verified / LiveCodeBench numbers do not exist yet; you have the Arena result and the Artificial Analysis Index of 57 (#4/189), and that is all.
• Planning to self-host? Wait for July 27 and confirm the weights actually ship under the promised license.
• Just want to test it cheaply? Run your own harness through the gateway today and compare against your current model on your tasks, not a vendor's table.

Kimi K3 is a real and impressive release. The honest summary is narrower than the headlines: two independent signals point up, the per-suite coding numbers are still Moonshot's own, the model is frontier-priced, and the “open” label is a date on a calendar, not a download link — yet.

FAQ