Kimi K3: What 2.8 Trillion Parameters Actually Buys You

Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter Mixture-of-Experts model that activates only ~50 billion parameters per token, making it the largest open-weight model ever built. K3 leads open-weight models on long-horizon coding benchmarks like SWE Marathon (42.0 vs. GLM-5.2's 13.0) but trails proprietary models on general intelligence, scoring 57 on the Artificial Analysis Intelligence Index v4.1 behind Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9). The model weights will be available on Hugging Face on July 27 under a modified MIT license, requiring approximately 1.4 TB of GPU memory for self-hosting.

Moonshot AI dropped Kimi K3 on July 16 and the benchmarking crowd immediately went into hyperdrive. Biggest open-weight model ever built. 2.8 trillion parameters. Beats Claude Fable 5 in the Frontend Code Arena https://www.kimi.com/blog/kimi-k3 . The model weights land on Hugging Face on July 27. Before you rewrite your infrastructure plans based on a benchmark leaderboard screenshot, here is what this actually changes for developers – and what it does not. The Architecture Behind the Number Calling K3 a “2.8 trillion parameter model” is technically accurate and practically misleading. Like DeepSeek and GLM before it, K3 is a Mixture-of-Experts architecture that activates only 16 of its 896 expert networks per token – roughly 50 billion equivalent active parameters per inference step. You are not paying to run 2.8 trillion parameters. You are paying to route through a 50B-equivalent model with exceptionally good routing. Three architectural decisions make K3 different from its predecessors. Kimi Delta Attention replaces quadratic self-attention in a subset of layers with a linear hybrid, enabling 6.3x faster decoding on million-token prompts. Attention Residuals let any layer selectively retrieve representations from earlier in the sequence, improving information retention across deep reasoning chains. Stable LatentMoE uses quantile-based expert routing that eliminates the load-balancing hyperparameters that make large MoE models brittle to tune. The model uses MXFP4 quantization applied during training, not as a post-training compression step. The result is roughly 1.4 TB storage for the full model versus the 5.6 TB an FP16 version would require. What the Benchmarks Actually Say K3 leads the open-weight field on the coding benchmarks that matter most for agentic workloads. On Terminal-Bench 2.1 it scores 88.3, half a point behind GPT-5.6 Sol for the best overall result. On SWE Marathon – the benchmark designed to measure autonomous coding capability over extended, multi-step horizons – K3 scores 42.0 compared to GLM-5.2’s 13.0. That is not a small gap. On Program Bench 77.5 and FrontierSWE 81.2 it is the best open-weight result available. On the Artificial Analysis Intelligence Index v4.1, K3 scores approximately 57, behind Claude Fable 5 59.9 and GPT-5.6 Sol 58.9 . Moonshot’s own announcement acknowledges K3 “still trails the most powerful proprietary models overall.” The lead K3 has is specifically in long-horizon autonomous coding – not in general intelligence benchmarks or conversational quality. The Three-Way Open Model Race Kimi K3 is not competing in isolation. Two strong open-weight models are available now, with weights already on Hugging Face under MIT licenses. DeepSeek V4 Pro is approximately 13 times cheaper per normalized token than K3. GLM-5.2 runs at nearly three times the inference speed. The full benchmark comparison https://www.marktechpost.com/2026/07/18/kimi-k3-vs-deepseek-v4-pro-vs-glm-5-2-open-trillion-scale-moe-models-compared-on-benchmarks-license-and-serving-cost/ from MarkTechPost breaks down each dimension in detail. | Kimi K3 | DeepSeek V4 Pro | GLM-5.2 | | |---|---|---|---| | Parameters | 2.8T total | 1.6T total | 744B total | | Active params | ~50B | ~49B | ~40B | | API Input | $3.00 per MTok | $0.435 per MTok | $1.40 per MTok | | API Output | $15.00 per MTok | $0.87 per MTok | $4.40 per MTok | | Cached Input | $0.30 per MTok | $0.004 per MTok | $0.26 per MTok | | Speed | ~62 tok/s | ~62 tok/s | ~168 tok/s | | Weights | July 27 | Available now | Available now | | License | Modified MIT | MIT | MIT | The Self-Hosting Reality: 1.4 TB of VRAM Every time a large open-weight model drops, someone asks if they can run it on a Mac Studio. The answer for K3 is no, and the bar is significantly higher than you might expect. The full model requires approximately 1.4 TB of GPU memory in MXFP4 format, before accounting for KV-cache and activations during a million-token context run. The practical minimum for experimental deployment is an 8-node cluster with 8 H100 80GB GPUs each – 5.12 TB total. That cluster costs roughly $1.6-2 million to purchase or approximately $40,000 per month to rent. Northflank’s self-hosting guide https://northflank.com/blog/what-is-kimi-k3-self-hosting lays out the full infrastructure stack once weights are available. Self-hosting K3 is a hyperscaler decision, not a startup infrastructure conversation. For the overwhelming majority of teams, the API is the only viable path. When K3 Is Worth It There are specific workloads where K3’s combination of long-context comprehension, strong agentic coding benchmarks, and frontier-class open weights creates genuine value: Long-horizon autonomous coding agents that run for minutes or hours across large repositories Full monorepo analysis where you need 500K+ tokens in active context, not just retrieval over a chunked index GPU kernel optimization, compiler development, and research automation Fine-tuning on proprietary data once the weights are available July 27 Where you should reconsider: K3 defaults to maximum thinking effort, which adds latency and output token cost to every request. At $15 per million output tokens, a verbose reasoning trace adds up fast. The Bigger Signal K3’s pricing tells you something important about where the open model ecosystem is heading. Its input cost is three to four times what K2.6 charged. Moonshot is not subsidizing frontier access to build an ecosystem – they are pricing frontier open models at frontier prices. The bifurcation between commodity open inference DeepSeek at $0.435 per MTok and frontier open inference K3 at $3.00 per MTok is now structural. The weights dropping on July 27 matter primarily for research teams and organizations with genuine fine-tuning requirements. K3 is the best open-weight model for long-horizon autonomous coding. It is not the cheapest, fastest, or most immediately available option – and for most tasks, those factors win.