API Update Overview
Kimi K3 is a 2.8-trillion-parameter open-source MoE model released by Moonshot AI on July 16, 2026, with immediate API availability (source). The world's first open 3T-class model, featuring native vision and a 1-million-token context window.
In official evaluations, K3 consistently outperforms all other tested models, though still trailing Claude Fable 5 and GPT 5.6 Sol.
Architectural Innovations
K3's architecture centers on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with quantization-aware training and fully-balanced expert parallelism for efficient inference (source).
Kimi Delta Attention (KDA)
KDA optimizes attention information flow across sequence length. A KDA prefix cache implementation has been contributed to vLLM, achieving >90% cache hit rate on long-context API workloads for dramatically lower inference cost.
Attention Residuals (AttnRes)
AttnRes selectively retrieves representations across depth rather than accumulating them uniformly, improving training stability in very deep models.
Stable LatentMoE
K3 scales MoE to 896 experts (16 active). Quantile Balancing derives allocation from router-score quantiles. Per-Head Muon optimizes attention heads independently.
Quantization & Inference Optimization
K3 applies quantization-aware training from SFT onward, using MXFP4 weights + MXFP8 activations for broad hardware compatibility. Fully-balanced expert-parallel training (static shapes, no host sync on critical path) prevents throughput degradation at large scales. 64+ accelerator supernode configurations are recommended for optimal inference.
Coding Capabilities
K3 excels at long-horizon autonomous coding — competing with Claude Fable 5 on kernel optimization, and building its own GPU compiler and chip (source).
Kernel Optimization
In head-to-head comparison with Claude Fable 5 and GPT 5.6 Sol, K3 completed profiling, rewriting, and benchmarking across NVIDIA Hopper GPUs and alternative GPGPUs within 24 hours, substantially outperforming Opus 4.8 and GPT 5.6 Sol.
GPU Compiler Development
K3 built MiniTriton from scratch — a Triton-like GPU compiler with its own tile-level IR, 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 (Nangate 45nm). Within 4 mm²: 1.46M standard cells, 0.277 MB SRAM, INT4 MAC array at 8,700+ tokens/s decode.
API Pricing & Availability
K3 is live on the Kimi API Platform (platform.kimi.ai) — model weights were open-sourced on July 27, 2026.
Pricing
| Item | Price | Notes |
|---|---|---|
| Cache-hit input | $0.30 / MTok | >90% hit rate on coding |
| Cache-miss input | $3.00 / MTok | First request or expired cache |
| Output | $15.00 / MTok | Per generated token |
With Mooncake's disaggregated inference and KDA prefix caching, cache hit rates exceed 90% on coding workloads.
Access
Visit the Kimi API Platform and select kimi-k3. 64+ accelerator supernode deployment is recommended.
Limitations
K3's main limitations are thinking history sensitivity and over-proactiveness — use explicit behavioral constraints in API system prompts (source).
- Thinking history sensitivity: K3 was trained in preserved thinking history mode. Agent harnesses must pass back all thinking content. Switching models mid-session may cause instability.
- Over-proactiveness: K3's training emphasizes long-horizon tasks. Provide more explicit behavioral instructions in API calls.
- Overall user experience still shows a gap vs Claude Fable 5 and GPT 5.6 Sol.