Fireship’s Code Report tested Kimi K3.1, a new model from Moonshot that scores close to the best paid coding models. If you use AI to build or maintain WordPress sites, plugins or scripts, you should know what is solid, what is self-reported, and what to test yourself before you switch tools.
What Kimi K3.1 is and how it scores
Moonshot calls it a 3-trillion-parameter open-weight model. Fireship says the weights should be fully downloadable on November 4th. Another source puts the size at 2.8 trillion total parameters, with about 104 billion active per token. The downloadable weights are due by November 4, 2026. The hosted version is available now.
On Terminal Bench 4.0, Moonshot reports 65.7 for K3.1 Pro. Claude Opus 5.5 scores 66.4 and GPT-6 Astra scores 58.2. Fireship says K3.1 is within one point of Opus, and that no other open model comes close.
On SWE Marathon, a test of multi-hour software jobs, Moonshot reports 51.7. Opus scores 46.3 and GPT-6 Astra scores 42.5. This test is most relevant if your work involves long, multi-step changes rather than small snippets.
The context window is 1 million tokens, so you can feed in a large plugin or theme codebase.
How far to trust the numbers
Most of these figures come from Moonshot. The source says some of the evaluations are internal and not independently verified. Fireship notes that independent testing is more reassuring. Artificial Analysis ran the model and got 67% on Terminal Bench, the highest of any model it tested. Another source confirms the 67% figure and ranks the model fourth of 215 on its Intelligence Index, with a score of 55.
Fireship also warns that the test harness changes the results a lot. Running different models with different harnesses can produce big differences. So a leaderboard number won’t tell you how the model behaves in your own editor or agent setup.
What the horse test shows
Fireship’s informal test asks a model for a realistic, interactive 3D horse in Three.js, then scores the result from 0 to 10. Google Argon scored 0.69, GPT-6 Astra scored 5.21, Opus scored 6.39, and Kimi K3.1 scored 4.20.
Kimi’s horse had the best colouring and the most realistic look. The animation was poor, though. Kimi moved the horse by warping the original mesh. Claude built a mesh of 95,000 vertices tied to 35 bones, each animated on its own. That is why Claude’s horse doesn’t look torn apart when it walks.
Limits and what it needs
Moonshot deliberately removed cybersecurity material from its training. It took cyber-related data out of two training stages. On its internal exploit suite, an unfiltered checkpoint solved 90.5% of tasks, while the released model solves 81%. These results are internal and have not been independently reproduced. If you want an AI helper for security research or penetration testing, expect it to be restricted by design.
Fireship’s view is that this is a sensible business move. The US recently banned China’s DeepSeek model from government use, including all federal civilian agencies, and Moonshot has little choice but to restrict its hacking ability. If you work with government clients, check your own rules on Chinese models before you adopt one.
It also doesn’t win everywhere. In Moonshot’s own table, K3.1 Pro scored 31.5 on Program Bench, against 65.0 for Opus. On Terminal-Bench-Science 0.1 it scored 38.1, against 63.3 for Opus. Don’t assume it beats paid models across the board.
Running it yourself is a serious hardware job. The raw weights are about 5.6 TB at BF16, 2.8 TB at 8-bit, 1.4 TB at 4-bit and 700 GB at 2-bit, before memory for the KV cache and overhead. Four 80 GB GPUs (320 GB) fall well short even of the 4-bit estimate. For most small teams, the hosted API is the realistic route.
The sources I read did not name the licence, so check it when the weights ship.
Cost matters too. Fireship tried the free tier, was denied, and paid 250 yuan per month for the demo. Standard API pricing is $2 per million input tokens and $10 per million output tokens. The faster tier costs $4 and $20, with a claimed two-to-three-times speed gain.
Your next step
Pick one real task from your WordPress work, such as a small plugin fix or a theme template change. Run the same prompt through Kimi K3.1 and the model you use today. Compare how each one structures the code, not only how the output looks.
Do this before November 4 on the hosted API, using a staging site. Then decide whether the price and the results justify changing tools. Read the licence once the weights are published.
