MLog
Back to posts
产品动态#AI智能体#开源#Kimi K3#大模型

Kimi K3: A 2.8T Open Model That Can Code, Design Chips, and Edit Videos

Published: Jul 18, 2026Reading time: 4 min

On July 17, Moonshot AI unveiled Kimi K3 — the world's first open 3T-class model. With 2.8T parameters, KDA attention, and a 1M-token context window, it goes head-to-head with proprietary giants across coding and agentic benchmarks.

In the early hours of July 17, 2026, Moonshot AI dropped Kimi K3 — a 2.8-trillion-parameter model with a 1-million-token context window and native vision capabilities. It is the world's first open 3T-class model, and it doesn't just raise the bar for open-source AI — it puts it squarely in the ring with GPT-5.6 Sol and Claude Fable 5.

Elon Musk commented "Impressive." Arena AI's Frontend Code Arena ranked it #1 with a score of 1679.

More Than Just Big Numbers

The 2.8T parameters grab headlines — for nine of the past twelve months, Kimi models have set the upper bound of open-model sizes. But the real story is in the architecture.

K3 is built on two key innovations:

  • Kimi Delta Attention (KDA): A hybrid linear attention mechanism that improves information flow across sequence length
  • Attention Residuals (AttnRes): Selectively retrieves representations across depth rather than accumulating them uniformly

On the MoE side, K3 activates 16 of 896 experts, paired with a Stable LatentMoE framework for robust routing. According to Moonshot, these structural changes deliver an approximate 2.5× improvement in scaling efficiency over Kimi K2 — more intelligence per unit of compute.

Training incorporates quantization-aware techniques from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility.

Coding: A New Benchmark for Long-Horizon Tasks

Coding is where Kimi K3 shines brightest.

Benchmark Kimi K3 Claude Fable 5 GPT-5.6 Sol
Terminal Bench 2.1 88.3 84.6 88.8
SWE Marathon 42.0 35.0 39.0
Program Bench 77.8 76.8 77.6
DeepSWE 67.5 70.0 73.0
FrontierSWE 81.2 86.6 71.3

K3 ranks first on SWE Marathon, the benchmark for ultra-long-horizon software engineering. On Terminal Bench 2.1 it edges past Fable 5. On real-world repository fixes (DeepSWE, FrontierSWE) it still trails the proprietary leaders — but the gap is narrowing fast.

The case studies are even more telling:

  • Chip design: In a single 48-hour autonomous run, K3 designed, optimized, and verified a complete chip using open-source EDA tools on the Nangate 45nm library
  • GPU compiler: Built MiniTriton from scratch — a compact Triton-like compiler with its own tile-level IR, optimization passes, and PTX codegen pipeline
  • Research reproduction: Completed in ~2 hours what would take an experienced researcher 1-2 weeks — a computational astrophysics pipeline involving 20+ papers, 300+ equations of state, and 3,000+ lines of Python

Agents: Search, Operate, Edit

On agentic benchmarks, K3 holds its own:

Benchmark Kimi K3 Claude Fable 5 GPT-5.6 Sol
BrowseComp 91.2 88.0 90.4
Automation Bench 30.8 29.1 29.7
SpreadsheetBench 2 34.8 33.4 33.7

Its multimodal chops are equally impressive. K3 achieves true "vision in the loop" — iterating between code and live screenshots during game development and frontend work. Its own 36-second teaser video was autonomously edited from 56 source clips, handling clip selection, motion-matched cuts, and frame-accurate beat synchronization.

Open Source and Availability

Full model weights will be released on July 27, 2026. Until then, K3 is available through:

  • Kimi.com and the Kimi mobile app (iOS / Android / HarmonyOS)
  • Kimi Work desktop (Windows / Apple Silicon Mac)
  • Kimi Code terminal tool
  • Kimi API platform

API pricing: $0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, $15.00/MTok for output. Powered by Mooncake's disaggregated inference architecture, the official API achieves over 90% cache hit rate on coding workloads.

Limitations and Outlook

Moonshot is upfront about K3's weaknesses: sensitivity to thinking history (switching models mid-session degrades quality), occasional over-proactiveness on ambiguous tasks, and a still-noticeable experience gap versus Fable 5 and GPT-5.6 Sol.

But the signal is unmistakable. Anthropic's CEO once predicted that an open model matching Fable 5 would arrive in 6-12 months. Kimi K3 has arrived earlier — and harder — than anyone expected.