Kimi K3's open weights landed with a #1 coding benchmark score — and a hallucination rate nobody put in the chart
Moonshot AI's Kimi K3 went open-weight on 27 July 2026, a 2.8-trillion-parameter model that ranks #1 on the Frontend Code Arena and scores 76.8% on SWE-bench Verified, but independent testing from Artificial Analysis found a hallucination rate of roughly 51% — up from 39% on the prior Kimi K2.6 — a figure Moonshot's own benchmark release omitted.
28 July 2026
Kimi K3’s open weights went live at 00:00 UTC on 27 July 2026 — free to download, 2.8 trillion parameters, a 1-million-token context window, and coding benchmark results that would headline any release: 76.8% on SWE-bench Verified, first place on Frontend Code Arena, and top rankings across Program Bench, SWE Marathon and several other agentic coding evaluations. On the numbers Moonshot AI chose to publish, it’s the strongest open-weight coding model released to date, and the largest open-weight release the industry has seen.
The number Moonshot didn’t put on its own benchmark chart is the one that matters most for anyone weighing whether to use it: Artificial Analysis’s independent AA-Omniscience testing measured K3’s hallucination rate at approximately 51%, up from 39% on its predecessor Kimi K2.6. That’s not a small regression dressed up as a tradeoff — it’s a majority-of-the-time failure rate on the specific measure of whether the model’s outputs are actually true, sitting directly alongside the industry’s best-scoring open coding model on task completion.
This is exactly the split this site has been tracking since the “AI code vulnerability trust gap” and “AI coding trust gap” posts earlier this month: benchmark leaderboards measure whether a model can produce plausible, working-looking code on a defined task set, not whether what it tells you about that code — the edge cases it claims to have handled, the security assumptions it states as fact — is true. A model that tops SWE-bench and hallucinates half the time isn’t a contradiction; it’s a model that’s very good at looking finished.
So what
If a lower-cost open-weight model with frontier coding benchmarks is on your shortlist for a build — your own team’s or a vendor’s — the benchmark score is not the diligence question. The diligence question is what verification sits between the model’s output and what ships: code review discipline, test coverage, and someone who actually reads what the model claims rather than trusting the confidence of the output. That verification layer is the difference between a cheap model being a genuine cost saving and being an expensive mistake discovered in production. It’s also the core of what AI-assisted development means done properly — using these tools for speed without inheriting their failure modes. If you’re evaluating an AI-heavy build and want that reviewed properly, get in touch.