Kimi K3: Moonshot AI's 2.8-Trillion-Parameter Open Model Edges Past GPT-5.6 on Coding Benchmarks
Source: arXiv, Kimi Team / Moonshot AI — 2026-07-27
Summary
Moonshot AI released the technical report for Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with 104B activated parameters per forward pass, native vision, and a 1-million-token context window. On reported benchmarks it scores 76.8% on SWE-bench Verified, 88.3 on Terminal-Bench 2.1, and 77.8 on ProgramBench — edging out GPT-5.6 and matching or beating Claude/Fable-class models on some coding tasks — while shipping as an open-weight model.
Key Takeaways
- Kimi K3 is a Mixture-of-Experts (MoE) model: 2.8T total parameters, but only 104B are "activated" per token — the model routes each request through a small subset of specialized sub-networks rather than running the full parameter count every time, which is what makes a model this large computationally feasible to serve.
- Reported scores: 76.8% on SWE-bench Verified (real-world GitHub issue resolution), 88.3 on Terminal-Bench 2.1 (agentic terminal/shell tasks), 77.8 on ProgramBench — all self-reported by Moonshot, not yet independently reproduced.
- 1-million-token context window plus native vision support, positioning it as a general-purpose frontier competitor rather than a coding-specialist model.
- The headline claim is that it edges out GPT-5.6 on some of these coding benchmarks while being open-weight — a meaningfully different distribution model than closed frontier labs, if the numbers hold up under independent testing.
Reel Script
Hook An open-weight model just posted coding benchmark scores that beat GPT-5.6 on paper — and you can download the weights yourself and check.
Core Concept Kimi K3 is what's called a Mixture-of-Experts model, or MoE. Instead of one giant neural network that processes every request in full, imagine a hospital with 2.8 trillion dollars' worth of specialists on staff, but for any given patient you only call in the roughly 104 billion dollars' worth of specialists actually relevant to that case. That's the trick that lets a model with 2.8 trillion total parameters actually run at a reasonable cost — you get the capacity of a huge model without paying to compute through all of it on every single token. Pair that with a million-token context window and native vision, and you get a model built to be a general-purpose competitor, not a narrow coding tool.
Hands-On The numbers Moonshot reported: 76.8% on SWE-bench Verified — that's the benchmark where a model has to actually resolve real GitHub issues, not just write plausible-looking code. 88.3 on Terminal-Bench 2.1, which tests whether an agent can actually operate a shell and get multi-step tasks done, not just answer questions about one. And 77.8 on ProgramBench. Moonshot's own framing is that these edge out GPT-5.6 on some of these tasks. The important caveat: these are Moonshot's self-reported numbers from their own technical report — nobody outside the company has independently reproduced them yet, so treat "beats GPT-5.6" as a claim to verify, not a settled fact.
Takeaway If even half of these numbers hold up independently, Kimi K3 is a serious argument that open-weight models are closing the gap with closed frontier labs on agentic coding tasks specifically — not just general chat quality. Don't take the benchmark table at face value; if you're evaluating coding models for your own stack, run your own eval on your own codebase before trusting anyone's self-reported leaderboard, including this one.