On the evening of July 17, Kimi released K3 — 2.8 trillion parameters, million-token context, the world’s first open-source 3T-class model.
Unsurprisingly, English and Chinese media reached for the same narrative frames: “domestic model chases the closed giants,” “another DeepSeek moment,” “K3 tops Arena, overtakes Fable 5.”
The reports are not wrong. They miss what matters. If Kimi K3 were a routine parameter upgrade, this analysis would not exist. K3 is something else: beneath the benchmark rankings sits a structural change — AI capability competition is switching from “what can a model answer” to “what can a model execute.” This is not a model launch. It is a fork in the road.
Parameters Are the Entry Fee, and Benchmark Rankings Are Losing Meaning
Start with the most easily missed fact: K3’s 2.8T parameters are not the story. The open parameter race has run for over a year — K2’s 1T to K3’s 2.8T — and the industry is numb to parameter inflation. What matters is which architectural innovations made that scale usable. Two deserve attention: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). KDA lets information flow across long sequences and deep stacks; AttnRes fixes attention’s information degradation in deep networks. With the Stable LatentMoE framework, K3 activates 16 of 896 experts and improves scaling efficiency roughly 2.5× over K2.
In plain terms: they found the engineering method to grow the model without exploding inference cost. That matters far more than rankings, because it answers an open question the community had left dangling — must open models accept being a generation behind? K3’s answer: no, provided you innovate hard on architecture rather than piling on compute.
The benchmarks back it up with a nuance the headlines flattened: K3 topped Arena’s frontend coding test over Fable 5, scored 74.7 on Vals AI’s composite (second only to Fable 5, ahead of GPT-5.6), and ranked first on SWE Marathon’s ultra-long coding tasks and BrowseComp’s deep web research. But Kimi’s own site concedes K3 still trails Fable 5 and GPT-5.6 Sol on general white-collar work close to real office flows. So K3 is not a straight-A student — it proves open models can reach world class in specific high-value domains. That “partial surpassing” is more informative than parity: model competition is becoming multidimensional and scenario-based, not one-dimensional.
What K3 Actually Changed: From Able to Chat to Able to Work
If you only read parameters and ranks, you miss the most interesting part. Kimi demonstrated K3 autonomously producing a 42-year research site on the ASIC industry: not an article, not an answer — 120+ self-directed iterations, 2,800+ web searches and scrapes, 1,100+ terminal data extractions, processing 87 quarterly reports and 99 raw PDFs into a research narrative with custom charts, animations, and interactive visualizations. Again: no human drove this.
Another case: in a continuous 48-hour agent run, K3 independently completed a chip’s construction, optimization, and verification on open-source EDA tools and a 45nm library. And there was the da Vinci workshop simulator that stunned overseas developers — K3 generated not just the visual style but understood the history of flapping-wing machines, their materials, and aerodynamics, weaving all of it into the interactive experience.
The common thread: it is not that K3 knows more — it is that K3 can do more. Traditional LLM evaluations (MMLU, HumanEval, GSM8K) measure what a model knows. K3’s cases measure what it completes — a different dimension entirely: knowledge reserve versus execution capability. K3 sustains these tasks because it was designed for long-horizon agent-style execution — chaining programming, visual understanding, tool calls, and knowledge work into complete flows rather than single question-answer rounds. One task can span hundreds of iterations with cross-tool dispatch and self-correction. That is why K3 loses some classic benchmarks to Fable 5 yet competes on complex task execution: the existing benchmarks do not test end-to-end completion at all.
A Two-Layer Model for Reading AI’s Capability Map
K3’s lesson is that we need a new frame for capability evolution — call it the two-layer model of AI capability.
Layer one: conversational intelligence. The focus of the past three years. Metrics: knowledge reserve, reasoning accuracy, language fluency; evaluated by Q&A, multiple choice, code generation. The layer’s ceiling: correctly answering a hard question — one-directional, passive, single-turn.
Layer two: execution intelligence. The dimension K3 is defining. Metrics: autonomous planning, multi-step execution robustness, tool-call accuracy, error recovery, deliverable completeness. The ceiling: starting from a fuzzy goal and autonomously completing an end-to-end task — multi-turn, proactive, delivering finished work.
The layers stack rather than replace. Layer one is the foundation — you cannot do well what you do not know. But excellence in layer one does not guarantee layer two; in fact, many MMLU stars collapse sharply on agent tasks beyond ten steps. The frame explains K3’s unique position precisely: still behind Fable 5 and GPT-5.6 Sol on conversational intelligence, already world-class on execution intelligence — leading in some scenarios like deep web research and long-horizon coding.
It also explains the pricing: 20 yuan per million input tokens and 100 per million output — above K2.6’s 6.5/27, far below Fable 5’s $10/$50 (roughly 68/339 yuan). Kimi is repricing from “talking rates” to “working rates”: a model that executes complex tasks occupies a different value dimension.
What the Frame Predicts
The two-layer model reaches beyond K3: Why did Anthropic ship Cowork? Because Claude’s conversational intelligence nears its ceiling; growth lives in the execution market. Cowork is Anthropic’s first productized move into layer two. Why did OpenAI ship Operator and Codex Agent? Same logic — GPT-5.6’s conversation is good enough; the headroom is downstream in execution. Why can open models beat closed ones in specific execution scenarios? Because execution intelligence optimizes for different objectives. A team obsessed with “keep the model running in a terminal for 48 hours without crashing” can out-execute a team optimizing conversation quality, even with slightly weaker chat. K3’s KDA architecture is built exactly for that: long sequences, multi-step, sustainable execution. Why is competition among Cursor, Windsurf, and peers intensifying? They are bridges from conversational to execution intelligence — and the stronger models’ native execution gets, the less bridging the tools must do, changing the tools’ competitive paradigm.
The frame also reads the price trend: if conversational intelligence is an information good and execution intelligence is a service good, the former keeps racing toward zero under open-source pressure while the latter stratifies by delivered value. K3’s pricing already hints at it.
What Each Role Should Do
Developers and tech leads: start testing K3’s agent mode now, especially long-horizon coding and terminal operations. Its SWE Marathon and Terminal Bench results suggest it may beat your current setup on “run it all day” workloads — and after July 27’s open release, private deployment and fine-tuning become worth the effort.
Product managers and founders: re-examine your product architecture. If your moat is “the model can answer questions,” K3-class execution intelligence enables “the product completes tasks autonomously.” Not an increment — a category possibility. Ask which ten-plus-step workflows in your market — market research, competitive analysis, code audit — just became fully automatable.
Researchers and model engineers: study KDA, AttnRes, and Stable LatentMoE. Open source’s bottleneck was never parameter count — it is inference efficiency at scale, and K3 demonstrates the architectural path around it.
Advanced everyday users: give K3 a fuzzy goal instead of a specific question. Don’t ask it to “explain transformers” — ask it to “build an interactive webpage introducing how transformers work.” The first is conversational intelligence. The second is execution intelligence. K3’s strength is the second.
Core thesis: Kimi K3’s value lies not in parameter scale or benchmark rank but in showing, for the first time, an open model’s path from conversational intelligence to execution intelligence. The first half of the parameter race is over. The second half — task execution — has begun.
