Kimi K3's 2.8T Parameters Meet the Pelican Test

Moonshot AI announced Kimi K3 on July 16, 2026, calling it the Chinese lab’s “most capable model to date” and claiming 2.8 trillion parameters. The model is available through Moonshot’s website and API, with an open-weights release promised by July 27, 2026.

Moonshot describes K3 as the first “open 3T-class model,” apparently rounding 2.8 trillion parameters up to 3 trillion. That would put it ahead of DeepSeek’s 1.6T v4 Pro by parameter count. In self-reported benchmarks, K3 generally beats Claude Opus 4.8 max and GPT-5.5 high, but trails Claude Fable 5 and GPT-5.6 Sol.

Artificial Analysis reported an Elo score of 1547 on its private long-horizon knowledge-work evaluation—732 points higher than Kimi K2.6 and behind only Claude Fable 5.

“On our private long-horizon knowledge work evaluation, Kimi K3 reaches an overall Elo of 1547, +732 points from Kimi K2.6 and behind only Claude Fable 5.”

The model’s task cost is $0.94, close to GPT-5.6 Sol at $1.04, roughly half the $1.80 cost reported for Opus 4.8, and higher than open-weights competitors. K3 also used 21% fewer output tokens than K2.6 on the Artificial Analysis Intelligence Index. It currently leads Arena.ai’s Frontend Code arena, surpassing Claude Fable 5.

Kimi K3 pricing and token use

K3 costs $3 per million input tokens and $15 per million output tokens—the same general pricing tier as Anthropic’s Claude Sonnet series, but the highest price yet for a model from a Chinese AI lab. That is a substantial increase over Kimi K2.6, priced at $0.95/$4. K3's 2.8 trillion parameters also represent more than twice the size of that earlier 1T model.

Recommended reading

OpenAI Hardware and Critical Flaws Shake Up Tech

To test it without creating a Moonshot API key, the benchmark author used OpenRouter and the llm-openrouter plugin to generate an SVG of a pelican riding a bicycle:

llm -m openrouter/moonshotai/kimi-k3 'Generate an SVG of a pelican riding a bicycle'

The prompt consumed 95 input tokens and 16,658 output tokens, including 13,241 reasoning tokens, for a total cost of 25 cents. K3 then analyzed the rendered SVG for 0.6 cents, producing a detailed description of the white pelican, red scarf, red bicycle, road, sky, clouds, sun, birds, and grass.

What the pelican benchmark still shows

The SVG test is now 21 months old. It began as a joke about how difficult it is to compare models, then developed a surprisingly strong correlation with overall model quality during its first year. That relationship has weakened: the GPT-5.6 and Claude Fable 5 pelicans are reportedly outclassed by GLM-5.2, which the benchmark author does not consider a Fable-class model.

The test’s larger weakness is that it says little about the capabilities that matter most in current systems: agentic tool calling and reliable tool use across long conversations. It should not be treated as a serious standalone model comparison.

It remains useful as a hands-on check. Running one prompt confirms that a model can be accessed, provides a rough estimate of cost and reasoning, and tests whether it can produce valid SVG with basic geometry and spatial awareness. Those checks are especially useful for smaller models running locally through tools such as llama.cpp, LM Studio, or Ollama.

K3's result is a clear improvement over Kimi 2.5, while also revealing a current limitation: it offers only one reasoning setting, “max.” The model spent nearly four times as many tokens reasoning as it returned in its response—13,241 versus 3,417. A simple “hi” prompt also used 86 tokens, suggesting an approximately 85-token hidden system prompt, which K3 refused to reveal.

The benchmark remains a practical “hello world” for model access, prompting, SVG generation, and vision. It is also a tradition on Hacker News, where readers reportedly ask for the pelican whenever a new model is reviewed.