The one-line version: two Chinese labs shipped open-weight, trillion-parameter coding models within weeks of each other — Meituan's LongCat-2.0 and Moonshot's Kimi K3 — and the benchmark scores everyone is quoting to rank them aren't on the same test. Strip out the leaderboard theater and the decision for a solo founder is simple: LongCat-2.0 is the cheapest near-frontier open coder to run in a loop, MIT-licensed; Kimi K3 is the more capable generalist that also sees images, at 4–10× the token price.
The benchmark trap everyone is walking into#
Here is the sentence that should stop you: LongCat-2.0's headline coding score is 59.5 on SWE-bench Pro, and Kimi K3's is about 76.8% on SWE-bench Verified. Those look like a blowout for Kimi K3. They are not a comparison at all.
SWE-bench Pro and SWE-bench Verified are different benchmarks. Pro is the newer, harder set — the tasks are longer-horizon and the ceiling is lower, so a 59.5 on Pro can represent stronger real-world engineering than a higher number on the easier Verified set. Meituan is explicit that its 59.5 edges GPT-5.5's 58.6 on the same Pro test and calls overall quality comparable to Gemini 3.1 Pro (VentureBeat). Kimi K3's Verified and Terminal-Bench (88.3%) numbers are real and strong — but they were earned on other tests, and coding-agent scores swing 10–26 points just from changing the harness around the model (NxCode).
The one real idea in this piece: when two models are benchmarked on different tests, the leaderboard is noise. The signal that survives is economics and license — and that's where these two genuinely diverge. (If reading benchmarks skeptically is new to you, we walked through the number wars in how to read an agent-memory benchmark — the same reflexes apply here.)
What each one actually is#
LongCat-2.0 (Meituan) is a 1.6-trillion-parameter Mixture-of-Experts model that activates only about 48 billion parameters per token, using a "LongCat Sparse Attention" design and a native 1-million-token context. It's released under a plain MIT License. Its most telling credential isn't a benchmark: for roughly two months it ran anonymously on OpenRouter under the alias "Owl Alpha," climbing to the top of the coding leaderboards on pure user preference before Meituan revealed it (Decrypt). It was also trained and served end-to-end on domestic Chinese ASIC superpods — the supply-chain story we covered in LongCat-2.0 was trained on domestic chips.
Kimi K3 (Moonshot) is a 2.8-trillion-parameter model activating about 104 billion parameters per token via a Stable LatentMoE architecture (16 of 896 experts), with native vision and a 1M-token context. It's the frontier of open-weight generalists right now — it doesn't just write code, it reads screenshots and diagrams — and its open weights landed in late July 2026 (we tracked that drop in the Founder's Wire, and pitched it against a frontier closed model in Claude Opus 5 vs Kimi K3).
The economics — where the decision actually lives#
LongCat-2.0 activates half the parameters and lists at a fraction of the price. If your agent loops over a codebase all day, that gap is your bill.
On OpenRouter, LongCat-2.0 lists around $0.30 / $1.20 per million input/output tokens during its launch promotion and roughly $0.75 / $2.95 after it ends. Kimi K3 lists around $2.90 / $14. That's 4–10× cheaper on output for LongCat-2.0. Now stack the architecture on top: LongCat activates ~48B parameters per token to Kimi K3's ~104B, so it's also cheaper to serve if you run your own inference. For an agent that reads and edits code in a loop — the exact workload both models are built for — cost-per-completed-task is the number that decides your runway, not the leaderboard.
"Self-hosting" is a data-center decision, not a laptop one#
Neither of these is a model you run on a workstation. Both are trillion-plus-parameter MoE models. Even quantized to INT4, LongCat-2.0 sits in the ~1TB-of-VRAM class — a multi-node cluster. Kimi K3 ships in MXFP4 (it was quantization-aware trained from the SFT stage), which helps, but real production traffic still wants eight or more high-memory GPUs. For a team of one, "self-host" honestly means renting an 8×H200-class node by the hour — and for most people the cheap open-weight API is the rational default. If you truly need on-prem for data-residency or air-gap reasons, LongCat-2.0's lower activated-parameter count makes it the more tractable of the two to stand up. (If your target is a single-GPU coder, that's a different model class — see Qwen3-Coder-Next vs Kimi K3 on one GPU.)
The founder's decision, in three lines#
- Pick LongCat-2.0 if you want the cheapest near-frontier open coder to run in an agent loop, you don't need image input, and you value the clean MIT license. It's the cost-per-task winner.
- Pick Kimi K3 if you need a model that also sees — screenshots, diagrams, UI — or you want the higher published coding-agent scores and token price isn't your binding constraint.
- Route between them if you can: LongCat-2.0 for the high-volume code edits, Kimi K3 for the multimodal or hardest steps. That's the move that gets you frontier-adjacent quality on a solo budget.
One caveat worth repeating before you commit: read Kimi K3's exact license terms. "Open weights" is not automatically "MIT-permissive." LongCat-2.0's MIT release is the cleaner commercial posture; if the weights are load-bearing in your product, the license is a business decision, not a footnote. And whichever you pick, verify today's numbers against each model's first-party page — launch-window pricing and benchmark claims move fast.
