Moonshot AI releases Kimi K3 — a 2.8T open-weights MoE model with million-token context

On 2026-07-16, Moonshot AI introduced Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with 16 active experts out of 896 total, a 1-million-token context window, and native vision capability (Moonshot AI, 2026-07-16). Per Moonshot AI, it is the first openly available model at the 3T-parameter scale, with full weights promised by July 27. Kimi K3 trails proprietary frontier models Claude Fable 5 and GPT-5.6 Sol in overall capability, but outperforms every other tested model across coding, knowledge work, and agentic benchmarks.

What happened

Kimi K3 launches on Kimi.com, the Kimi Work desktop app (v3.1.0+), Kimi Code (via /model), and the Kimi API at platform.kimi.ai. Pricing: $0.30/MTok cache-hit input, $3.00/MTok cache-miss input, $15.00/MTok output. The Mooncake disaggregated inference architecture drives over 90% cache hit rates in coding workloads (Kimi API Platform).

The model uses max thinking effort by default; low- and high-effort modes will arrive in a later update. The full technical report is forthcoming alongside the weights release.

Why it matters

Kimi K3 is the first open model to reach the 2.8T-parameter scale — nine of the past twelve months, Kimi models have set the open-model size record. More importantly, it demonstrates that open-weight models can now compete with proprietary systems on agentic coding and long-horizon tasks. This has practical implications for teams that need model-level auditability, custom fine-tuning, or on-premise deployment.

Architecture highlights

Kimi K3’s major architectural components (Moonshot AI, 2026-07-16):

• Kimi Delta Attention (KDA) — linear-complexity attention enabling efficient inference at 1M-token context. The FlashKDA CUDA implementation is open-source under MIT license (FlashKDA).
• Attention Residuals (AttnRes) — selective depth-wise retrieval of representations across layers. Claims ~25% higher training efficiency at under 2% additional parameters. Published on arXiv (2603.15031).
• Stable LatentMoE — 16/896 expert activation with Quantile Balancing, which derives allocation from router-score quantiles rather than heuristic updates.
• Gated MLA — Multi-head Latent Attention variant with a gating mechanism for improved attention selectivity.
• SiTU activation — Sigmoid Tanh Unit replacing standard activation functions.
• Per-Head Muon optimizer — head-independent optimization extending the Muon algorithm.
• Quantization-aware training from the SFT stage: MXFP4 weights, MXFP8 activations.

The blog reports an approximate 2.5× scaling efficiency improvement over Kimi K2.

Coding and agentic performance

Kimi K3’s strongest results are in coding benchmarks. Key numbers (Moonshot AI, 2026-07-16):

In an autonomous chip design proof of concept, K3 produced a 45nm design at 100 MHz within 4 mm² in a single 48-hour run, using open-source EDA tools. It also built MiniTriton, a from-scratch GPU compiler that matches or beats Triton and torch.compile on roofline benchmarks (Moonshot AI, 2026-07-16).

Practical implications

• Open weights (by July 27) — teams can self-host, fine-tune, and audit the model, unlike GPT-5.6 Sol or Claude Fable 5.
• API cost — $3/MTok input and $15/MTok output is competitive for a model at this scale, helped by the >90% cache hit rate.
• Deployment requirements — the blog recommends supernode configurations with 64+ accelerators for inference, which is not a casual deployment but within reach for organizations with existing GPU clusters.
• Kimi Code integration — the model is available immediately via the Kimi Code CLI agent, making it practical for terminal-based coding workflows.

Risks and caveats

1. Weights are not yet released — the July 27 promise is credible given Moonshot’s track record (K2, K2.5, K2.6 were all released as Apache-2.0), but the actual license and availability are TBD.
2. No independent technical report yet — the blog provides high-level architecture descriptions but the detailed technical report is promised alongside the weights.
3. Known limitations (Moonshot AI, 2026-07-16):
   • Thinking history sensitivity — K3 requires preserved thinking history; harness compatibility matters.
   • Excessive proactiveness — may make unexpected decisions on ambiguous tasks; needs explicit behavioral constraints for bounded applications.
   • UX gap — the blog acknowledges a noticeable experience gap compared to Fable 5 and GPT-5.6 Sol.
4. Benchmark harness variation — several comparisons use different harnesses (KimiCode vs Claude Code vs Codex), which the blog’s footnotes disclose but may affect cross-model comparability.
5. Proprietary frontier gap — K3 “still trails” Fable 5 and GPT-5.6 Sol across the overall evaluation suite.

What to watch

1. Weights release on July 27 — the actual license terms, model card, and community reception will determine real-world adoption.
2. Technical report — expected alongside weights; will clarify training data, compute budget, and detailed evaluation methodology.
3. vLLM integration — Moonshot has contributed KDA prefill cache support to the vLLM community, which should simplify self-hosted deployment.
4. Ecosystem adoption — whether fine-tuned variants, quantizations (GGUF, AWQ), and toolchain integrations appear quickly.
5. Comparison to DeepSeek’s next model — DeepSeek V3/R1 series is the main open-weight competitor at this scale; the relative standings will shift as both release updates.