Let me tell you something I did not expect to write this week.
For years, when Western AI folks talked about Chinese models, the conversation went roughly like this: “They’re behind, they’re copying, they’re getting close.” Incremental. Reactive. A step back from the frontier.
Kimi K3 broke that script for me.
On July 16, 2026, Moonshot AI – a Beijing-based startup most Western developers had filed away as “the Kimi chatbot company” – dropped a model that not only set a world record for open-source parameter count but also climbed to the top of a coding benchmark that both Claude Sonnet 5 and GPT-5.6 Sol had been fighting over. First place. As an open-source release.
I have been running K3 through its paces for the past few days, reading every benchmark I could find, and talking to developers who got early access. Here is what actually matters – no fluff, no nationalism, just an honest look at what this model can do and what it means for the industry.
What K3 Actually Is – No Hype
Kimi K3 is a 2.8 trillion parameter Mixture-of-Experts model. That makes it the largest open-source model by parameter count in the world right now, about 75% bigger than DeepSeek V4 Pro. During inference, only 16 of its 896 expert modules activate per token – so while it “knows” a vast amount, you only pay for the compute used by a fraction of it.
The key technical advances Moonshot built into this thing:
Kimi Delta Attention (KDA) is a hybrid attention mechanism that replaces most standard Transformer attention computations with linear variants, keeping a small fraction of full-attention layers for precise retrieval tasks. The ratio is 3:1. This delivers a 75% reduction in KV cache usage and up to 6x faster decoding at 1M token context length compared to a standard architecture at the same context length.
Attention Residuals (AttnRes) is a mechanism that lets the model selectively retrieve information across layers – addressing the classic problem where deep models “forget” what earlier layers learned as the signal propagates upward. In plain terms: deeper models usually bury early-layer knowledge. AttnRes gives the model a way to reach back and grab it when needed.
Together, these two innovations produced roughly 2.5x the training efficiency of K2, according to Moonshot’s own benchmarks. The model also trained with MXFP4 quantized weights and MXFP8 activations from the supervised fine-tuning stage – meaning deployment efficiency was baked in from the start, not retrofitted.
Oh, and full open-source weights drop July 27, 2026. We are writing this 10 days before that date, so everything here is based on hosted access and published benchmarks.
The Benchmark That Actually Mattered
Here is the moment that made the AI community on X lose its collective mind:
Kimi K3 ranked first on the Frontend Code Arena (Arena.ai), beating both Claude Sonnet 5 and GPT-5.6 Sol. This is the first time an open-source model has topped this particular leaderboard. Let that sink in – open-source, not closed, not a lab with a decade of infrastructure, a Beijing startup.
The scores that back this up:
– Terminal-Bench 2.1 (terminal + agent workflow): K3 88.3% vs GPT-5.6 Sol 88.8% vs Sonnet 5 84.6%
– Program Bench (multi-step programming): K3 77.8% vs GPT-5.6 Sol 77.6% vs Sonnet 5 76.8%
– AA-Briefcase (long-context agent work): K3 1527 vs GPT-5.6 Sol Max 1495
Where K3 still trails: complex engineering benchmarks like DeepSWE and FrontierSWE. Moonshot openly publishes this, which I respect – it is more honest than most Western labs manage.
What K3 Actually Built – The Real Test
Benchmarks are warm-up exercises. Here is what K3 produced in its public demonstrations that actually got engineers talking:
A complete chip in 48 hours. K3 designed, optimized, and verified a chip from scratch using open-source EDA tools and the Nangate 45nm process library. 4mm², 1.46 million standard cells, timing convergence at 100MHz, simulated decoding throughput over 8,700 tokens/second. That is not a toy. That is a real piece of silicon designed by an AI.
A compiler built from scratch. K3 developed MiniTriton, a Triton-like compiler with its own tile-level MLIR intermediate representation, complete optimization passes, and PTX code generation. On Roofline benchmarks it outperformed Triton and torch.compile on several workloads. They even trained nanoGPT with it – and it converged normally.
Two hours of astrophysics research. K3 read 20+ papers, evaluated 300+ equation-of-state models, and generated 3,000+ lines of code to reproduce an I-Love-Q universal relation study. A senior researcher typically charges one to two weeks for that work.
A 42-year ASIC industry report generated through 120+ rounds of recursive self-improvement, 2,800+ web searches, 1,100+ terminal data pulls, processing 87 quarterly reports and 99 PDFs totaling 11,000+ pages. Inside Kimi Work, no less.
The common thread across all of these: long-horizon autonomy. K3 ran these tasks for hours at a time without losing the thread. That is the real capability upgrade.
How to Get Your Hands on It
K3 is live right now. Here is the practical rundown:
Kimi Website and App
The fastest path. Open kimi.com or the Kimi mobile app. K3 is the default model as of the July 16 launch. No configuration needed.
Kimi Code (CLI)
For developers who live in the terminal, Kimi Code is the most powerful way to use K3. Here is the exact step-by-step:
Step 1 – Install (one line):
curl -fsSL https://kimi.moonshot.cn/install | bashStep 2 – Start:
kimiStep 3 – Authenticate via browser:
/loginThis opens your default browser for a one-click login. No API keys needed for the hosted version.
Step 4 – Switch to K3:
/model k3After switching, the context window indicator will show 1M – confirming you have the full million-token window active.
Step 5 – Verify installed skills:
/skillKimi Code automatically scans and loads skills from the standard ~/.agents/skills/ directory – the same location Claude Code and other agent frameworks use. If you already have skills set up for another AI coding tool, they load automatically here with no extra configuration.
The skills I tested with K3 include:
– Context7 – live documentation search for any library or framework
– Firecrawl – web search and full-page scraping with markdown output
– browser-use – autonomous browser control for end-to-end testing
Once authenticated and switched to K3, you are ready to start projects. The model defaults to max thinking intensity, which triggers the deepest reasoning chains. Low and High modes are on the roadmap for a future update.
Official API
curl -X POST https://api.moonshot.cn/v1/chat/completions \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-d '{
"model": "kimi-k3",
"messages": [{"role": "user", "content": "Your prompt here"}],
"temperature": 0.3
}'Pricing is direct-to-Sonnet 5 competition:
– Input (cache hit): ~$0.28 per million tokens
– Input (cache miss): ~$2.75 per million tokens
– Output: ~$13.75 per million tokens
Moonshot claims 90%+ cache hit rate in programming scenarios via their Mooncake split-personality inference architecture, bringing effective input costs to roughly $0.07/M tokens. We will verify this independently after full weights drop.
Kimi Work
The knowledge worker desktop product with integrated tool use – web search, PDF parsing, code execution, and recursive self-improvement loops. This is where the ASIC report demo ran.
What I Actually Think After Using It
Here is my honest take, not a press release:
What impressed me most is the multimodal self-correction loop. K3 can screenshot its own output, identify visual errors, and fix them autonomously. For developers building web applications, this is genuinely new – you are essentially giving the model a pair of eyes to check its own work. That changes how you prompt.
The 1M token context is real and the KV cache reduction from KDA makes it more practical than raw token counts suggest. I ran it against a 400-page legal document and a 50,000-line legacy codebase simultaneously – it held the full context without the latency cliff I have hit on other “long context” models.
On code specifically – the Frontend Code Arena result is legitimate, but my experience aligns with the benchmark gap: K3 is exceptional at frontend and scripting work, strong at terminal workflows, and still slightly behind on very large-scale refactoring or complex multi-repo engineering tasks. Know where to use it.
The thinking mode defaults to “max” which means more internal reasoning steps before responding. For quick queries this can feel slow. Low and High modes are coming – I would hold judgment on latency until those land.
On pricing – the $0.28/M cache-hit rate is genuinely competitive for a model of this capability level. But the 90% cache hit claim is benchmarked on programming scenarios. For general conversation, expect lower hit rates and higher effective costs.
The Honest Limitations
Moonshot published three explicit limitations in their release notes. I will not pretend they are not there:
1. Full open-source weights not yet released (July 27). Everything we are evaluating right now is based on hosted access. The research community will validate the architecture claims soon.
2. Complex engineering benchmarks still trail Sonnet 5 and GPT-5.6 Sol. If your team is doing large-scale system refactoring across multiple repositories, K3 is not yet your best option.
3. The 90% cache hit rate is programming-scenario-dependent. Do not price your production budget assuming that number unless your use case mirrors programming workflows.
Why This Moment Matters
Here is the broader thing I keep thinking about.
For years, the narrative around Chinese AI development was “two years behind the frontier.” K3 does not erase that gap entirely – Sonnet 5 and GPT-5.6 Sol are still ahead overall – but it does something more interesting: it has claimed a specific, high-value territory (coding + agent autonomy) and declared it open-source.
The chip design demonstration is the thing I cannot stop thinking about. Not because most readers are chip designers, but because of what it represents. K3 did not just write code – it managed a 48-hour engineering project with multiple tool calls, failure recovery, and verification loops. That is a different category of capability than “generate a response to my prompt.”
I am not here to tell you K3 is better than Claude or GPT-5.6 Sol. It is not, overall. But it is good enough, on coding-specific tasks, at open-source prices, with a July 27 full release coming. That is a different kind of competitive pressure than anyone expected from a Beijing startup six months ago.
The frontier just got more crowded.
