<- all tokdocs
2026.07.17@chase_ai_

Kimi K3: China's 2.8 Trillion Parameter Open-Weight Model Tops the Frontend Code Arena

AI
Watch on TikTok

View on TikTok ->

Chase AI covers the July 16, 2026 release of Kimi K3, Moonshot AI's new flagship model. His argument in 53 seconds: Kimi K3 is a 2.8 trillion parameter Chinese open-weight model that beats Claude Fable 5 and GPT-5.6 in certain categories, but it burns far more tokens and runs slower, so the low per-token price is not the whole cost story.

The headline claim: #1 in the Frontend Code Arena

The opening frames show an Arena.ai leaderboard titled "Frontend Code Arena, Kimi-K3: Ranked #1." Kimi-K3 sits at the top with 1,679 points, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol (xHigh) at 1,618. GLM-5.2 (Max), Claude Opus 4.8 (Thinking), and Grok-4.5 fill out the next spots. Arena.ai confirmed the result on X and noted the jump: Kimi-K2.6 ranked #18 on the same board, so K3 climbed 17 places in one generation. Arena rankings come from Elo-style human voting on real coding tasks, which makes this a preference benchmark rather than a pass/fail test suite.

What the other benchmarks show

The video's middle section displays a benchmark chart with three coding evaluations, and the numbers on screen match Moonshot's published figures:

  • DeepSWE: GPT-5.6 Sol 73.0, Fable 5 70.0, Kimi K3 67.5, GPT-5.5 67.0, Opus-4.8 59.0, GLM-5.2 46.2
  • FrontierSWE: Fable 5 86.6, Kimi K3 81.2, GPT-5.6 Sol 71.3, GLM-5.2 67.3, Opus-4.8 66.7, GPT-5.5 64.9
  • Kimi Code Bench 2.0 (Internal): Fable 5 76.9, Kimi K3 72.9, Opus-4.8 71.7, GPT-5.5 69.0, GPT-5.6 Sol 64.8

Chase describes this accurately as "holding its own" rather than winning. K3 places second or third on these suites, behind Fable 5 on both FrontierSWE and Moonshot's own internal benchmark. He also flags that Kimi Code Bench 2.0 is Moonshot's internal benchmark, which is the right caveat, and even there the company shows itself losing to Fable 5.

The model itself

The second half of the video scrolls through Moonshot's official Kimi K3 docs page, which states the specs on screen: 2.8 trillion parameters, a hybrid linear attention mechanism (Kimi Delta Attention), Attention Residuals, and a 1M-token context window aimed at software engineering, knowledge work, and deep reasoning. It is a sparse mixture-of-experts model, activating 16 of 896 experts per token, which is how a 2.8T model stays affordable to serve. The docs page shows a Playground, API key signup, and an OpenAI SDK-compatible quickstart.

The scale doubles the roughly 1 trillion parameters of Kimi K2.6 and makes K3 the largest open-weight model announced to date. One timing note: at release the model was available through the app and API, with the weights download scheduled for July 27 under a modified MIT license. So the video's "open weight and open source" framing was a promise, not yet a download link, on the day it was posted.

The catch: token appetite and speed

Chase's caution is the most useful part of the video. Kimi K3's list price is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30. That undercuts Western flagship list prices, but K3 spends many more tokens per task. Artificial Analysis measured it generating 130M tokens to complete its Intelligence Index evaluation, against a 63M average for comparable models, and clocked output at 62 tokens per second, below the average for its class. A model that is twice as verbose can erase a per-token discount, and the slower throughput compounds on long agentic runs.

His hardware point holds up too. Open weights do not mean laptop-friendly. Moonshot recommends supernode configurations with 64 or more accelerators for self-hosting, so nearly everyone will consume K3 through an API.

Key takeaways

  • Kimi K3 took #1 on Arena.ai's Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618), a 17-place jump from Kimi K2.6.
  • On agentic coding suites it trails the leaders: behind Fable 5 on FrontierSWE (81.2 vs 86.6) and behind GPT-5.6 Sol and Fable 5 on DeepSWE (67.5 vs 73.0 and 70.0).
  • It is the largest open-weight model announced: 2.8T parameter sparse MoE, Kimi Delta Attention, 1M-token context, with weights scheduled for July 27 under a modified MIT license.
  • Cheap per token is not cheap per task: $3/$15 per million tokens, but measured verbosity of roughly double the class average and 62 tokens per second output.
  • Self-hosting is impractical for most: Moonshot recommends 64 or more accelerators, so API access is the realistic path.

Resources

Published July 17, 2026. Writeup generated from a favorited TikTok.