Chinese lab Moonshot AI announced Kimi K3 today, describing it as their "most capable model to date, with 2.8 trillion parameters." The model is available now via Moonshot's website and API, and the company has pledged to release open weights by July 27, 2026.
Positioning and benchmarks
Moonshot is billing Kimi K3 as the first "open 3T-class model" (the 2.8 trillion parameter count is effectively rounded up in their messaging). That puts it ahead of DeepSeek's 1.6T v4 Pro in terms of scale.
According to self-reported benchmarks from Moonshot and results reported by Artificial Analysis, Kimi K3 generally outperforms Anthropic's Claude Opus 4.8 max and OpenAI's GPT-5.5 high on many tasks, while trailing Claude Fable 5 and GPT-5.6 Sol.
Highlights from the Artificial Analysis report include:
- "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."
- "Cost per task ($0.94) is similar to GPT-5.6 Sol ($1.04), ~1/2 the price of Opus 4.8 ($1.80) and higher than open weights peers."
- "Kimi K3’s token usage on the Artificial Analysis Intelligence Index decreased significantly, using 21% fewer output tokens than K2.6."
Separately, Kimi K3 is currently the top model on Arena.ai's Frontend Code arena, even surpassing Claude Fable 5 in that leaderboard.
Pricing
Moonshot charges $3 per million input tokens and $15 per million output tokens for Kimi K3. This price point aligns it with Anthropic's Claude Sonnet series and makes it the most expensive model released so far by a Chinese AI lab. For comparison, the earlier Kimi K2.6 was priced at $0.95/$4. The 2.8T parameter size is also more than double the scale of the ~1T models Moonshot released previously.
A practical test: the pelican-on-a-bicycle prompt
To get hands-on experience, the author ran a familiar toy benchmark — "Generate an SVG of a pelican riding a bicycle" — via OpenRouter using the llm-openrouter plugin to avoid creating a Moonshot API key. The SVG generation consumed 95 input tokens and 16,658 output tokens, of which 13,241 were reasoning tokens, resulting in a cost of about $0.25.
Because Kimi K3 accepts image input, the author also sent the rendered SVG back with an alt-text prompt; K3 returned a detailed alt-text description for roughly $0.006.
The generated alt text described a cartoon-style white pelican with a red scarf riding a red bicycle on a gray road, with a light blue sky, sun, small black birds in flight, and green grass with tiny white flowers in the foreground.
What the pelican test reveals
The author notes several takeaways derived from this quick experiment:
- Kimi K3 currently exposes only one reasoning effort level, "max," and that shows in high reasoning-token consumption — the pelican example used a large number of reasoning tokens, which made the query relatively expensive.
- Tokenization behavior is noteworthy: a short prompt like "hi" counted as 86 tokens on Kimi K3 in the author's measurements, suggesting a hidden system prompt of roughly 85 tokens that the model declined to reveal.
- Vision capabilities appear strong: the automatic alt-text produced was high quality.
The author cautions that the pelican prompt began as a lighthearted benchmark and has lost some of its correlation to actual model quality over time. Importantly, the test does not probe agentic tool use and long-horizon tool reliability — capabilities that are increasingly central to modern model evaluation.
Still, the pelican test serves practical purposes: it forces an actual API run, provides a rough cost and reasoning-token estimate for a simple task, and confirms the model's ability to produce valid SVG and basic spatial reasoning. The Kimi K3 pelican output is a clear improvement over Kimi 2.5.
Conclusion
Moonshot AI's Kimi K3, at 2.8 trillion parameters and with a planned open-weight release on July 27, 2026, represents a significant step in large open models. Reported benchmarks show strong performance relative to several leading models, although some closed models remain ahead on particular metrics. The model's relatively high pricing and substantial reasoning-token usage will be important factors for users evaluating cost and practical utility.
