DEV Community

Cover image for Kimi K3: Moonshot AI's 2.8-Trillion-Parameter Open Frontier Model — Benchmarks, Architecture, and Everything We Know
AgentOne profile image BestCodes
BestCodes for AgentOne

Posted on • Originally published at agent-one.dev

Kimi K3: Moonshot AI's 2.8-Trillion-Parameter Open Frontier Model — Benchmarks, Architecture, and Everything We Know

#ai #machinelearning #opensource #llm

Heads up: This article was written by AgentOne Research.

Moonshot AI has officially launched Kimi K3, a 2.8-trillion-parameter mixture-of-experts (MoE) model that the Beijing-based startup is billing as the world's first open "3T-class" AI system. With a 1-million-token context window, native vision capabilities, and a new architecture built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), K3 is being positioned as a direct challenger to Claude Fable 5 and GPT-5.6 Sol — at roughly half the price.1

Full model weights are promised by July 27, 2026. Until then, developers can access K3 via kimi.com, the Kimi API (accessible via AgentOne and the Kimi Platform), Kimi Code, and Kimi Work.

The Specs at a Glance

Specification Detail
Total Parameters 2.8 trillion
Active Experts 16 of 896 routed experts (Stable LatentMoE)
Context Window 1,048,576 tokens (1M)
Input Modalities Text, images, video
Output Modality Text
Architecture KDA + AttnRes + Stable LatentMoE + Gated MLA
Training Format Quantization-aware training from SFT; MXFP4 weights, MXFP8 activations
Default Max Output 131,072 tokens (configurable up to context limit)
Reasoning Effort Max only at launch; low/high coming later
API Model ID kimi-k3
Open Weights Promised by July 27, 2026
Recommended Hardware Supernode with 64+ accelerators

Architecture: How Moonshot Scaled to 2.8T

Kimi K3 is not just a bigger K2. It is built on two architectural innovations developed internally at Moonshot AI, plus a scaled-up MoE sparsity framework.2

Kimi Delta Attention (KDA)

KDA is a hybrid linear attention mechanism that interleaves linear-attention layers with periodic full-attention layers in a 3:1 ratio. Three linear layers handle local sequence structure cheaply, while one full-attention layer preserves global information flow. According to Moonshot's research, this design cuts KV-cache memory by up to 75% and delivers up to 6× faster decoding at 1M-token contexts — all while matching or beating full-attention baselines on short-context, long-context, and post-training tasks.3

Attention Residuals (AttnRes)

AttnRes replaces standard residual connections with a mechanism that selectively retrieves representations across model depth rather than accumulating them uniformly. Moonshot reports this delivers roughly 25% higher training efficiency at under 2% additional cost.4

Stable LatentMoE

K3 activates 16 out of 896 experts per token. To handle the imbalance this creates, Moonshot introduced:

  • Quantile Balancing — derives expert allocation directly from router-score quantiles, eliminating heuristic updates and sensitive balancing hyperparameters.
  • Per-Head Muon — extends the Muon optimizer by optimizing attention heads independently.
  • Sigmoid Tanh Unit (SiTU) and Gated MLA — improve activation control and attention selectivity.

Together, these advances yield an approximate 2.5× improvement in overall scaling efficiency compared to Kimi K2.5

Inference Optimizations

K3 uses quantization-aware training from the SFT stage onward, with MXFP4 weights and MXFP8 activations for broad hardware compatibility. Because KDA poses new challenges for prefix caching, Moonshot contributed an implementation to the vLLM community.6

Benchmark Numbers

All Kimi K3 results below use reasoning effort set to max, temperature = 1.0, and top-p = 1.0. Depending on the benchmark, models are evaluated under one of three agentic harnesses: KimiCode, Claude Code, or Codex.7

Coding Benchmarks

Coding Benchmarks

Kimi K3 leads on Program Bench (77.8) and SWE Marathon (42.0), comes very close to GPT-5.6 Sol on Terminal-Bench 2.1 (88.3 vs 88.8), and trails Fable 5 on FrontierSWE (81.2 vs 86.6) and DeepSWE (67.5 vs 70.0).8

Agentic & Productivity Benchmarks

Agentic Benchmarks

On agentic tasks, K3 leads on BrowseComp (91.2), Automation Bench (30.8), and SpreadsheetBench 2 (34.8), while trailing Fable 5 on GDPval-AA v2, AA-Briefcase, and JobBench.9

Multimodal & Knowledge Benchmarks

Knowledge Benchmarks

K3 is competitive across knowledge and multimodal tasks, leading on MathVision (97.8) and OmniDocBench (91.1), while trailing slightly on CharXiv Reasoning and MMMU-Pro.10

Independent Ranking: Artificial Analysis v4.1

Artificial Analysis independently evaluated Kimi K3 Max and placed it #4 by configuration and effectively #3 by model family on its Intelligence Index v4.1.11

AA v4.1 Ranking

K3's 0.54-point gap from Sol xhigh is smaller than Artificial Analysis's estimated 95% confidence interval of roughly one point, suggesting the difference may not be statistically significant. The 1.78-point gap from Sol Max and 2.75-point gap from the Fable configuration are clearer, though workload-specific testing is still needed.

Cost and Token Use Across the Top Three

Configuration AA v4.1 Output Tokens (index) Total Eval Cost
Claude Fable 5 Max, Opus 4.8 fallback 59.9 87 million $5,630.52
GPT-5.6 Sol Max 58.9 70 million $2,824.00
Kimi K3 Max 57.1 130 million $2,690.80

K3 used roughly 1.9× more output tokens than Sol and 1.5× more than Fable to complete the evaluation, yet its lower token prices kept the total bill slightly below Sol and far below Fable.12

API Pricing

Pricing Comparison

Moonshot reports a cache-hit rate above 90% on coding workloads with the official API. At list price, K3 input is 40% cheaper than GPT-5.6 Sol and its output is 50% cheaper. Against Claude Fable 5, both rates are 70% lower.13

Case Studies: What K3 Can Actually Build

1. Chip Design in 48 Hours

In a single autonomous run, K3 designed a physical chip to run a nano-scale version of itself. Using open-source EDA tools on the Nangate 45nm library, K3 completed the full construction pipeline — from architectural design through optimization and verification — in just 48 hours.14

The result: a 4 mm² chip that closes timing at 100 MHz, sustains over 8,700 tokens/s decode throughput in simulation, packs 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization.

2. MiniTriton: A GPU Compiler Built from Scratch

K3 developed MiniTriton, a compact Triton-like compiler with its own tile-level IR layer over MLIR, optimization passes, and a PTX code-generation pipeline. Across supported roofline benchmarks, MiniTriton delivers performance on par with or better than Triton and torch.compile — beating Triton on certain workloads. It also sustains end-to-end nanoGPT training with stable convergence.15

3. 3D Open World Game Development

Kimi K3 built a fully procedural browser-based 3D exploration game using Three.js WebGPU and GPU compute. It procedurally generated the environment with forests, a log-cabin village, snowy mountains, and dynamic weather, while using 3D asset generation tools for character models.

4. Computational Astrophysics Research

Kimi K3 reproduced the universal I-Love-Q relation in computational astrophysics in approximately two hours — work that typically takes a senior researcher one to two weeks. It reviewed and cross-validated 20+ papers, implemented the full numerical pipeline, evaluated 300+ equations of state, identified inconsistencies in published formulas, generated 3,000+ lines of Python code, and produced an interactive HTML dashboard.16

Knowledge Work & Agentic Capabilities

Beyond public benchmarks, Moonshot reports consistent gains across internal evaluations derived from real-world user-agent workflows.17 Kimi K3 in Kimi Work can produce interactive research reports with bespoke charts, animated diagrams, and visual narratives. For example:

  • 42 years of AI ASIC industry research: Created through 120+ rounds of recursive self-improvement, pulling data via 2.8k+ web searches/fetches and 1.1k+ terminal data pulls across 11k+ pages.
  • Fusion Industry Research: A consulting-style report with timelines, funnel charts, range bar charts, Gantt charts, and publication-quality slides.
  • GWTC-5 Gravitational-wave Analysis: Analysis of 391 events using 20+ concurrent subagents, producing 7 scientific visualizations, 2 tables, and literature synthesis from 10+ papers.

Kimi K3 also created a 3Blue1Brown-style motion-graphics explainer of its own architecture, translating technical ideas into animated diagrams and transitions. It edited its own teaser video from 56 source clips, handling clip selection, motion-matched cuts, frame-accurate beat synchronization, audio processing, and multiple rounds of revision.18

How K3 Ranks Against the Frontier

In tests of real-world task automation, Kimi K3 ranked first in four out of eight benchmarks — including Automation Bench, SpreadsheetBench 2, and BrowseComp — while finishing second to Fable 5 in most others. Fable 5 and GPT-5.6 Sol were its closest competitors overall.19

Kimi K3 also claimed the No. 1 spot on Arena.ai's Frontend Code Arena with a score of 1,679, outpacing Claude Fable 5 and GPT-5.6 Sol.20

The Open-Source Frontier Timeline

For nine of the past twelve months (July 2025 – July 2026), Kimi models have maintained the upper bound of open-model sizes. Kimi K3 continues that trajectory at 2.8T parameters, nearly triple the size of DeepSeek V4 Pro (1.6T) and more than double its immediate predecessor Kimi K2.6 (1T).21

Availability & Access

Platform Status
Kimi.com Live — sign up with Google or phone number
Kimi Work Desktop app (Windows, Apple silicon Mac), version 3.1.0+
Kimi Code Terminal agent — select K3 via /model command
Kimi API kimi-k3 on api.moonshot.ai
Open Weights Promised by July 27, 2026

Kimi K3 is compatible with the OpenAI SDK, lowering the integration barrier for developers already building on OpenAI or Anthropic toolchains. It supports streaming with separate reasoning_content and final-answer content deltas, structured JSON output, tool calling with dynamic loading, vision inputs, and a partial mode for prefix continuation.22

Limitations

Moonshot openly acknowledges three key limitations:23

  1. Sensitivity to thinking history. K3 was trained in preserved thinking history mode. If an agent harness fails to pass back all historical thinking content, generation quality may become highly unstable. Use a verified harness like Kimi Code and avoid switching to K3 mid-session.
  2. Excessive proactiveness. K3's training emphasizes long-horizon, challenging tasks. It may make unexpected decisions on the user's behalf when encountering minor issues or ambiguous intent. Impose explicit behavioral constraints in system prompts or AGENTS.md if your application requires strict boundaries.
  3. User experience gap. Despite being highly competitive overall, K3 still exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT-5.6 Sol.

The Bottom Line

Kimi K3 is the most significant open-weight release since DeepSeek V4 Pro, and it closes much of the performance gap with the leading closed-source frontier models. Its 57.1 score on Artificial Analysis v4.1 puts it effectively third among all model families, trailing only Claude and GPT-5.6 Sol. It beats both on several individual benchmarks — including BrowseComp (91.2), Program Bench (77.8), and Automation Bench (30.8) — while offering 40–70% lower token prices than its closest competitors.24

The catch: weights are not yet available, and the model's always-on thinking mode means high output-token consumption that can erode list-price savings. But if Moonshot delivers on its July 27 promise, K3 will reset the open-weight performance ceiling and give enterprises a credible, downloadable alternative to proprietary models.

For developers, researchers, and enterprises watching the open-source AI movement, Kimi K3 is the moment open-source stopped trailing by months and started trading blows at the frontier.

Footnotes

  1. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  2. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  3. Moonshot AI, Kimi Delta Attention paper (arXiv 2510.26692) and K3 technical blog. https://www.kimi.com/blog/kimi-k3 ↩

  4. Moonshot AI, Attention Residuals paper (arXiv 2603.15031) and K3 technical blog. https://www.kimi.com/blog/kimi-k3 ↩

  5. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  6. Moonshot AI, "Kimi K3 - Kimi API Platform," documentation, July 16, 2026. https://platform.kimi.ai/docs/guide/kimi-k3-quickstart ↩

  7. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  8. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026; DeepSWE leaderboard https://deepswe.datacurve.ai/; Program Bench https://www.vals.ai/benchmarks/programbench. https://www.kimi.com/blog/kimi-k3 ↩

  9. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  10. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  11. Artificial Analysis, "Kimi K3 Evaluation," Intelligence Index v4.1, July 16, 2026. https://artificialanalysis.ai/models/kimi-k3 ↩

  12. Artificial Analysis, "Kimi K3 Evaluation," Intelligence Index v4.1, July 16, 2026. https://artificialanalysis.ai/models/kimi-k3 ↩

  13. Moonshot AI, "Kimi K3 - Kimi API Platform," documentation and pricing, July 16, 2026. https://platform.kimi.ai/docs/guide/kimi-k3-quickstart ↩

  14. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  15. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  16. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  17. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  18. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  19. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  20. Arena.ai, Frontend Code Arena leaderboard, July 16, 2026. Cited in Moonshot AI blog and VentureBeat coverage. https://www.kimi.com/blog/kimi-k3 ↩

  21. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  22. Moonshot AI, "Kimi K3 - Kimi API Platform," documentation, July 16, 2026. https://platform.kimi.ai/docs/guide/kimi-k3-quickstart ↩

  23. Moonshot AI, "Kimi K3: Open Frontier Intelligence," official tech blog, July 16, 2026. https://www.kimi.com/blog/kimi-k3 ↩

  24. Artificial Analysis, "Kimi K3 Evaluation," Intelligence Index v4.1, July 16, 2026; Kingy AI analysis. https://artificialanalysis.ai/models/kimi-k3 ↩

Top comments (0)