Kimi K3, and what we can still learn from the pelican benchmark
- ID
- 5366
- Status
- summarized
- Published
- 17 Jul 2026, 4:19 AM
- Fetched
- 17 Jul 2026, 5:10 AM
- Provider
- Simon Willison
- Category
- developer-ai
- Original URL
- https://simonwillison.net/2026/Jul/16/kimi-k3/
- Source URL
- https://simonwillison.net/atom/everything/
Summary
- Score
- 8.0
- Created
- 17 Jul 2026, 5:10 AM
- Tags
- Audience
- developersvibe_codersai_ml_learnersai_agent_users
What happened
Simon Willison reviews Moonshot AI's new Kimi K3 model, highlighting its 2.8 trillion parameters, competitive benchmarks against Claude and GPT, and relatively high API pricing of $3/$15 per million tokens. He also tests the model using his informal 'pelican benchmark' for SVG generation, noting that the correlation between this test and overall model quality has weakened.
Why it matters
For Malaysian developers and AI builders, Kimi K3 represents another powerful option in the rapidly evolving LLM landscape, particularly for frontend code generation where it currently leads. The promised open-weight release by July 27, 2026, could offer local startups a path to self-hosting a state-of-the-art model, though the API pricing is higher than previous Chinese lab releases.
Discussion angle
How informal benchmarks like the 'pelican test' highlight the difficulty of evaluating modern LLMs, and whether Kimi K3's higher pricing is justified for production use compared to established alternatives.