Kimi K3 from Moonshot AI: what changes vs K2 and when to use it (2026)

July 23, 20267 minKimi K3, Moonshot AI, AI, LLM, Open Source

Short answer (60 seconds): Kimi K3 is the first open-source 2.8T-parameter model (Moonshot AI, weights released July 27, 2026), with a 1M-token window, native vision, and a new architecture (Kimi Delta Attention + Attention Residuals). API pricing: USD 0.30/MTok cache-hit input, USD 3.00 cache-miss, USD 15.00 output. Shines at: long-horizon coding, kernel optimization, agentic knowledge work, and multimodal with video. Choose it when: you need 1M context without compaction, want self-hosting post-release, or have workflows that benefit from > 90% cache-hit in coding. Don't choose it when: you prioritize polished UX and a mature ecosystem (Claude Fable 5 and GPT-5.6 Sol still win).

Kimi K3 is Moonshot AI's latest release and the first in their line to cross the 2T-parameter threshold with open weights. This post is the guide I wish I'd had on launch day: architecture, pricing, honest benchmarks, and the real question — what work is it better than Claude Fable 5 or GPT-5.6 Sol for, and what work isn't it?

The architecture: why KDA + AttnRes matter

What's new in K3 isn't just size — it's two architectural decisions that change how the model scales.

Kimi Delta Attention (KDA) is an attention variant that improves how information flows across long sequences. Combined with Attention Residuals (AttnRes), which selectively retrieves representations across layers instead of accumulating them uniformly, Moonshot reports ~2.5× improvement in scaling efficiency vs K2: the same amount of compute produces more intelligence measured by benchmarks.

K3 also scales Mixture-of-Experts sparsity: it uses Stable LatentMoE with 16 of 896 experts active per token. Routing is the first-order challenge at this scale — that's why Moonshot introduces Quantile Balancing (derives expert allocation from router-score quantiles, eliminating the sensitive balancing hyperparameter) and Per-Head Muon (optimizes attention heads independently for more adaptive learning at scale).

Other relevant components: SiTU (Sigmoid Tanh Unit) for activation control, Gated MLA for attention selectivity, and MXFP4 weights + MXFP8 activations from the SFT stage for broad hardware compatibility.

All of this sounds like buzzwords until you see the numbers: in long-horizon coding, Moonshot reports that K3 handles extended engineering sessions with minimal oversight, navigates large repos and orchestrates terminal tools stably. In the official post they describe how an early K3 build handled the majority of the team's kernel optimization work in late development stages.

Specs and pricing

Official API pricing:

For coding workloads, Moonshot reports cache hit rate above 90% — in practice, if your workflow has repeated prefixes (system prompt, common context), you pay input at the cache-hit price most of the time.

Benchmarks: where K3 is frontier and where it isn't yet

Moonshot compares K3 against Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, GPT-5.5, and GLM-5.2. The honest read:

K3 is competitive or wins at:

• Long-horizon coding (DeepSWE, Terminal-Bench 2.1, SWE Marathon, FrontierSWE)
• Kernel optimization (H200 and GPGPU alternate vendor)
• Agentic knowledge work (OfficeQA Pro, SpreadsheetBench 2, MCP Atlas, AutomationBench)
• Multimodal with video (frame-accurate editing, motion graphics)
• Reasoning over large repos with mixed code and visuals

K3 still doesn't reach Claude Fable 5 or GPT-5.6 Sol at:

• Polished user experience (Moonshot is explicit about this gap)
• Mature third-party tooling ecosystem
• Consistent latency in production at massive scale (2.8T weights weigh on inference)

In coding, the current ranking per published benchmarks positions K3 behind Fable 5 but ahead of Opus 4.8 in agent-driven tasks. On OfficeQA Pro, SpreadsheetBench 2, and GDPval-AA, the top three are very close and the choice depends more on ecosystem than pure benchmark.

Cases where K3 shines with real examples

The Moonshot blog shows public cases, not isolated benchmarks. Three worth understanding:

1. MiniTriton — a GPU compiler from scratch. K3 built a Triton-style compiler with its own tile-level IR over MLIR, optimization passes, and a PTX codegen pipeline. The output reaches performance comparable to Triton and torch.compile, and trains nanoGPT end-to-end with stable convergence. It's the difference between "optimizing isolated kernels" and "building a coherent compiler" — the latter requires many-hour agency.

2. Chip design in 48 hours. As proof of concept, K3 designed, optimized, and verified a chip to serve a nano model built on its own architecture, using open-source EDA tools over Nangate 45nm. It closed timing at 100 MHz in under 4 mm², with 8,700 tokens/s decode throughput in simulation.

3. Reproducible astrophysics research. To reproduce the I-Love-Q universal relations, K3 reviewed 20+ papers, implemented the numerical pipeline, evaluated 300+ equations of state, identified inconsistencies in published formulas, generated 3,000+ lines of Python, and produced an interactive HTML dashboard. What takes 1-2 weeks for an experienced researcher, K3 did in 2 hours.

These cases are cherry-picked by Moonshot, yes — but they're the kind of tasks where a founder with a SaaS isn't going to spend USD 50K on consulting to automate. K3 changes the economic calculation.

How to access Kimi K3

Four official paths:

• Kimi.com — web and mobile chat (iOS, Android, HarmonyOS)
• Kimi Work desktop app — version 3.1.0+ for Windows and Apple silicon Mac; ideal for agentic knowledge work
• Kimi Code — CLI for terminal with /model kimi-k3 command
• Kimi API Platform — kimi-k3 as model ID; pricing as in the table above
• Kimi Enterprise — separated data privacy, member management for teams

If you're going self-hosted: wait until July 27. Moonshot contributes the KDA implementation for prefix caching to the vLLM project, so you'll want vLLM with that patch to serve K3 at a competitive token price.

Limitations Moonshot admits explicitly

Three things worth keeping in mind before choosing K3 over alternatives:

1. Sensitivity to thinking history. K3 was trained in preserved thinking history mode. If your agent harness doesn't pass all thinking content back, or if you switch to K3 mid-session from another model, quality becomes very unstable. Solution: use a harness with verified compatibility (Kimi Code) and avoid switching mid-session.

2. Excess proactivity. K3's training emphasizes long-horizon, challenging tasks. Facing ambiguity or minor issues, it may make unexpected decisions on the user's behalf. If your application needs the agent to operate within well-defined boundaries, impose explicit constraints in the system prompt or in AGENTS.md.

3. UX gap. Moonshot is explicit: "a noticeable gap in user experience compared with Claude Fable 5 and GPT-5.6 Sol". The API works, but product polish and third-party ecosystem maturity aren't at Anthropic or OpenAI's level yet.

When to pick K3 — and when not

Pick K3 if:

• You need 1M tokens of context without compaction (BrowseComp score 90.4 with full context)
• You'll self-host post July 27 (lower total cost than API)
• Your workflow is long-horizon coding with repeated prefix (you pay cache-hit > 90% of the time)
• You work on research or agentic knowledge work where extended autonomy pays
• You want native vision integrated with text and video in the same model

Pick Claude Fable 5 or GPT-5.6 Sol if:

• You prioritize UX polish and mature tools/integrations ecosystem
• You do short coding with standard tool calling (most SaaS CRUD with AI)
• You need mature compliance (HIPAA, SOC2, etc.) — top providers have years of lead
• Your workload is sensitive to consistent latency

Still pick K2 if:

• You already have K2 in production and the upgrade isn't urgent
• Your inference hardware doesn't reach the 64+ accelerators Moonshot recommends for K3

Want to discuss how Kimi K3 fits your stack before spending USD 50K on implementation? There's a CTA at the end with a free 30-minute call.

Frequently asked questions

What is Kimi K3 in one sentence?

It's Moonshot AI's latest model: 2.8T parameters with Kimi Delta Attention + Attention Residuals architecture, 1M-token window, native vision, and multimodal reasoning. It's the first open-source model to cross the 2T-parameter threshold and competes with proprietary models on long-horizon coding and knowledge work.

How does it compare to Kimi K2?

K3 is ~2.5× more scaling-efficient than K2 (the same compute produces more intelligence measured by benchmarks), supports 1M context (vs ~200K for K2), and has native vision (K2 was primarily text). K3 uses Mixture-of-Experts with 16 of 896 experts activated per token (Stable LatentMoE).

Is Kimi K3 free?

Full weights are released on July 27, 2026 (open source). Until then, you can use it via paid API: USD 0.30/MTok cache-hit input, USD 3.00/MTok cache-miss input, USD 15.00/MTok output. For coding workloads, Moonshot reports cache hit rate above 90%, which in practice lowers the bill significantly.

What tasks does Kimi K3 shine at?

Long-horizon coding (multi-hour engineering sessions with minimal human oversight, navigation of large repos, terminal orchestration), kernel optimization, agentic knowledge work, research with interactive visualization, and multimodal with video (frame-accurate synchronized editing). In the official blog they show cases where K3 designed a chip in 48 hours or compiled 3000 lines of astrophysics research code in 2 hours.

When does Kimi K3 beat Claude Fable 5 or GPT-5.6 Sol?

For tasks that need 1M tokens of context without compaction, or when you want self-hosting (weights released July 27). For short coding with mature tool calling, Claude and GPT still win on user experience and ecosystem. Moonshot is explicit: K3 has a noticeable UX gap compared to top proprietary models.

Is Kimi K3 available in Spanish?

Yes, via API and at kimi.com. The model was trained with multilingual data (the internal benchmark measures performance in Chinese, English, and other languages). For Spanish-specific benchmarks I don't have public data at the time of this post — I'll update with numbers when Moonshot publishes the full version of the technical report.