Moonshot AI Ships Kimi K3, the Largest Open-Weight Model Ever at 2.8 Trillion Parameters

Moonshot AI released Kimi K3 on July 16, and the headline number is real: 2.8 trillion parameters, making it the largest open-weight model ever shipped. But the parameter count is the least interesting thing about it.

The architecture is a sparse Mixture-of-Experts design. According to the Kimi technical blog, the model introduces two new mechanisms: Kimi Delta Attention, a hybrid linear attention approach, and Attention Residuals. The practical result of that sparsity, as documented in Moonshot's own disclosures, is that only about 50 to 60 billion parameters actually activate on any given token pass. The 2.8 trillion figure reflects total weights, not compute per inference step. That distinction matters because it's what makes a model this large economically viable to serve.

Benchmark placement landed fast. According to a detailed breakdown by researcher Ken Huang, within hours of release Kimi K3 debuted at number one on LMArena's Frontend Code Arena, beating Anthropic's Claude Fable 5 on that specific leaderboard. Tom's Hardware, reviewing Moonshot's own disclosures, reported that K3 placed ahead of Claude Opus 4.8 and GPT-5.5 on coding and agentic benchmarks, while still trailing Claude Fable 5 and OpenAI's GPT-5.6 Sol on overall performance. Kili Technology's evaluation flagged a 51 percent hallucination rate on its tests, a number that sits alongside those top coding scores and isn't easy to dismiss for anyone considering production deployment.

The architecture has a few specifics worth naming. According to the Kimi API platform documentation, the model uses MXFP4 weights with MXFP8 activations and requires a supernode with at least 64 accelerators for local serving. The 1-million-token context window is enabled by Kimi Delta Attention, which Kili Technology's analysis found produces up to 6.3x faster decoding at long contexts compared with standard attention, at a claimed 2 percent compute overhead. That's not a minor spec-sheet detail: long-context performance is what separates a model that can ingest an entire code repository from one that can't.

The open-weight release is the other thing to watch closely here. The model launched in Kimi products and through the API on July 16, but as of this writing, the full weights were not yet public. The Kimi technical blog committed to releasing them by July 27. Until those weights drop and inference partners can verify deployment, the open-source claim has an asterisk. As NxCode noted in its evaluation guide, developers can run the hosted model and inspect the open Kimi Code evaluation harness, but they cannot reproduce weight-level serving or check training claims against the actual parameters. Moonshot says a full technical report will accompany the weight release.

Pricing, according to Moonshot's published API rates, is $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens. For agentic coding workloads where the system prompt and codebase context are reused across many requests, that cache-hit rate is the number that controls real cost.

The competitive context is straightforward. According to Tom's Hardware's review of Moonshot's technical blog, K3 took the open-weight size crown from DeepSeek's 1.6 trillion parameter V4. The previous ceiling for open-weight models was roughly one trillion parameters. Moonshot nearly tripled that in one release cycle. Whether the coding leaderboard win holds under the kind of evaluation rigor that production engineering teams apply is a different question, and one the field will answer over the next several weeks as the weights circulate.

Sources cited:
- Kimi API Platform documentation (https://platform.kimi.ai/docs/guide/kimi-k3-quickstart)
- Kimi K3 technical blog (Moonshot AI) (https://www.kimi.com/blog/kimi-k3)
- Tom's Hardware (https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3)
- Ken Huang, Substack (https://kenhuangus.substack.com/p/demystifying-kimi-k3-how-chinas-28t)
- Kili Technology (https://kili-technology.com/blog/kimi-k3s-benchmarks-and-hallucinations----what-that-tells-us-about-ai-evaluation)
- NxCode evaluation guide (https://www.nxcode.io/resources/news/kimi-k3-benchmarks-coding-agent-evaluation-guide-2026)