Kimi K3: The Largest Open Model Ever (2.8T Params), and Why Almost No One Can Run It Locally
Skip to content
Subscribe
Dark
On July 16, 2026, Moonshot AI announced Kimi K3 , and called it the first open model to cross 2.8 trillion parameters . Independent testers put it near the top of the field, and the press ran with "largest open-source model ever." For a local-LLM site the news is genuine and the timing is awkward: as of the announcement the weights are not downloadable. Moonshot says the full weights land by July 27, 2026 , and until then Kimi K3 is an API and a chat window, not a file you can pull.<br>So this is a preview, not an owner report. Nobody outside Moonshot has run the weights, there is no technical report yet, and there are no community quants. We have not tested it. What we can do is lay out what Moonshot claims, what the one independent eval so far shows, and the part that matters most here: what a 2.8-trillion-parameter model would take to run at home. The short version is that the machine most people would reach for, a maxed-out Mac Studio, cannot hold it at any quant.<br>What Moonshot says Kimi K3 is<br>Straight from Moonshot's announcement and its API docs, all of it vendor-stated:
SpecKimi K3 (as stated by Moonshot)
Total parameters~2.8 trillion (Mixture-of-Experts)<br>Active per tokenNot disclosed. Only the routing is given: "16 of 896 experts" active ("Stable LatentMoE")<br>Context window1,000,000 tokens<br>ArchitectureKimi Delta Attention (a hybrid linear attention) plus Attention Residuals<br>ModalitiesNative vision (text + images + video in), always-on "max thinking"<br>PrecisionQuantization-aware training in MXFP4 weights / MXFP8 activations<br>LicenseNot stated for K3. K2 and K2.5 shipped Modified MIT, so that is the expectation, not a confirmed fact<br>Available nowAPI + web only (model id kimi-k3); weights "by July 27, 2026"
Two of those lines deserve a flag. The license is genuinely unknown : the announcement names none, and the "Modified MIT" label floating around comes from aggregator sites inferring it from the K2 family, not from Moonshot. And the active-parameter count is not published , which is the single most important number for anyone thinking about running it. More on why below.<br>One detail local runners should note: Moonshot says K3 was trained quantization-aware in MXFP4 , the same 4-bit block format OpenAI's gpt-oss models use. If that holds through the weight release, the "native" quant is 4-bit-class, which sets the floor on how small the download can get without extra damage.<br>The benchmarks: read them as claims<br>Because the weights are not out, every per-benchmark number Moonshot published is a vendor self-report that no one can reproduce yet. Moonshot's own table leads with agentic and coding results: it claims 91.2 on BrowseComp (which it frames as state-of-the-art at release), 93.5 on GPQA-Diamond , and 88.3 on Terminal-Bench 2.1 . Take those as the maker's numbers, not settled facts. Note also there is no SWE-bench Verified score for K3 from any credible source, so any figure you see claiming one is invented.<br>The number worth trusting is the independent one. Artificial Analysis, which runs its own evals through the API, scored Kimi K3 at 57 on its Intelligence Index and ranked it third to fourth overall: on par with Claude Opus 4.8 and GPT-5.5, a step behind the current leaders (Fable 5 and GPT-5.6 Sol). That directionally backs Moonshot's positioning. Separately, blind human-preference testing on frontend coding (relayed by Axios and Simon Willison) had developers preferring K3 over every leading US model on that one narrow task. Willison also ran his usual "pelican on a bicycle" SVG doodle and got a valid result, while cautioning it is not a real capability test.<br>The takeaway: a Chinese lab shipping an open-weight model that an independent index rates alongside Opus 4.8 is the actual story here, and it lands whether or not the individual self-reported cells hold up under scrutiny after July 27.<br>The one spec that decides local speed is missing<br>For a model you plan to run yourself, two numbers matter, and they do different jobs. Total parameters decide whether it fits in your memory. Active parameters decide how fast it decodes , because token generation is memory-bandwidth-bound: each token only reads the experts that fire, not all 2.8T weights. Our Mixture-of-Experts explainer walks through why.<br>Moonshot gave the routing ("16 of 896 experts") but not a per-expert dimension, so the active-parameter count in billions cannot be derived from what is public. The prior generation, Kimi K2, was 1T total with 32B active. If K3 kept a similar active budget, it could decode at a usable clip despite its size; if the active count grew, it would be slower. We do not know, and any specific active-parameter figure circulating right now is an aggregator's guess. That single missing number is why a reliable tok/s estimate is impossible today.<br>The hardware reality: nothing you own runs...