Model Overview
Kimi K3 is a 2.8-trillion-parameter open-source MoE model released by Moonshot AI on July 16, 2026 — the world's first to reach the 3T class (source). With native vision and a 1-million-token context window, K3 marks the entry of open-source models into the ultra-large-scale frontier.
In official evaluations, K3 consistently outperforms all other tested models, though it still trails the strongest proprietary models — Claude Fable 5 and GPT 5.6 Sol.
Architectural Innovations
K3's core innovations are Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), paired with Stable LatentMoE across 896 experts for stable large-scale training (source). Overall scaling efficiency improved ~2.5× over K2.
Kimi Delta Attention (KDA)
KDA optimizes attention information flow across sequence length for efficient scaling. A KDA-aware prefix caching implementation has been contributed to the vLLM community, enabling K3 to serve at highly competitive prices at long contexts.
Attention Residuals (AttnRes)
AttnRes selectively retrieves representations across depth rather than accumulating them uniformly, significantly improving training stability and information utilization in very deep models.
Stable LatentMoE
K3 scales MoE to 896 experts (16 active). Quantile Balancing derives expert allocation from router-score quantiles, eliminating heuristic updates and sensitive hyperparameters. Per-Head Muon optimizes attention heads independently.
Coding Capabilities
K3 excels at long-horizon autonomous coding tasks, from kernel optimization and compiler development to chip design (source).
Kernel Optimization
In head-to-head GPU kernel optimization against Claude Fable 5 and GPT 5.6 Sol, K3 completed profiling, rewriting, and benchmarking kernels across NVIDIA Hopper GPUs and alternative GPGPUs within 24 hours, competing strongly with Fable 5 and significantly outperforming Opus 4.8 and GPT 5.6 Sol.
GPU Compiler Development
K3 built MiniTriton from scratch — a compact Triton-like GPU compiler with its own tile-level IR layer, optimization passes, and PTX codegen pipeline. Roofline benchmarks match or beat Triton and torch.compile.
Chip Design
In a single 48-hour autonomous run, K3 designed a chip using open-source EDA tools on the Nangate 45nm library. Within 4 mm²: 1.46M standard cells, 0.277 MB SRAM, INT4 MAC array, sustaining 8,700+ tokens/s decode throughput.
Coding for Research
In a computational astrophysics test, K3 completed in about two hours what typically takes one to two weeks — reproducing I-Love-Q universal relations, cross-validating 20+ papers, evaluating 300+ equations of state, and generating 3,000+ lines of Python code.
Knowledge Work & Agents
K3 shows significant gains in end-to-end knowledge work, supporting interactive research, visual dashboards, and multimodal content creation (source).
- Interactive Industry Research: Produces drill-down reports through 120+ rounds of recursive self-improvement, extracting data from 2,800+ web searches
- Widgets & Dashboard: Kimi Work introduces interactive components and personalized dashboards
- Video Editing: K3's native multimodal architecture handles clip selection, beat synchronization, and multi-round revisions from 56 sources
Availability & Pricing
K3 is live across Kimi.com, Kimi Work, Kimi Code, and the Kimi API — weights were open-sourced on July 27, 2026 (API platform).
| Platform | How to access |
|---|---|
| Kimi.com | Use directly in browser |
| Kimi Work | Desktop app v3.1.0+ (Windows / Apple Silicon Mac) |
| Kimi Code | Use /model command in terminal |
| Kimi API | Select kimi-k3 model on the platform |
API Pricing
| Item | Price |
|---|---|
| Cache-hit input | $0.30 / MTok |
| Cache-miss input | $3.00 / MTok |
| Output | $15.00 / MTok |
With Mooncake's disaggregated inference architecture, cache hit rates exceed 90% on coding workloads.
Limitations
K3's main limitations are sensitivity to thinking history and excessive proactiveness — use explicit behavioral constraints in system prompts (source).
- Thinking history sensitivity: K3 was trained in preserved thinking history mode. Agent harnesses must pass back all historical thinking content. Switching models mid-session may cause quality instability.
- Excessive proactiveness: K3's training emphasizes long-horizon tasks. It may make unexpected decisions with minor issues or ambiguous intent.
- Overall user experience still shows a gap compared to Claude Fable 5 and GPT 5.6 Sol.