Kimi K2 is the best model ever (kind of…)

Detailed Summary

Introduction (0:00 - 0:50)

The video introduces Moonshot's Kimi K2 Thinking model as a significant new open-weight AI. It highlights the model's exceptional tool-calling capabilities, claiming it to be the best seen so far. Kimi K2 also achieves state-of-the-art scores on humanity's last exam and browser comp, and performs comparably to GPT-5 and Sonnet 4.5 in code benchmarks. The model's ability to perform 200-300 consecutive tool calls without human interference is emphasized, noting its "nuts" performance and the excitement surrounding its release.

Sponsor Segment: Tuple (1:05 - 2:52)

The video includes a sponsored segment for Tuple, a pair programming tool. The presenter demonstrates Tuple's features, such as shared control over a computer, drawing on screen, and real-time collaboration, emphasizing its utility for developers and its superiority over traditional screen sharing methods like Zoom and Slack.

Model Specifications and Hosting (2:52 - 4:57)

Kimi K2 Thinking is described as a massive model, both in its potential impact and its physical size. It boasts a trillion parameters and a 594GB size, making it the largest open-weight model ever created. While it is INT4 quantized to facilitate running, its sheer size still poses a challenge, with Moonshot currently being the primary host. On Open Router, Kimi K2 is available in two versions: a standard one at 60 cents/million input and 250/million output with 18 TPS, and a turbo version at $1.15/million input and $8/million output with 85 TPS. The pricing and speed are compared to GPT-5, noting similarities in their split offerings, suggesting Moonshot's ambition to rival OpenAI.

Benchmark Performance and Token Usage (4:57 - 7:12)

Kimi K2's benchmark results are presented as "nuts," leading all open-weight models according to Artificial Analysis's intelligence index. However, its speed is noted as "absolute garbage," and it is highly token-hungry, consuming 140 million tokens in the intelligence index, significantly more than GPT-5 (82 million) and Claude 4.5 Sonnet (34 million) for reasoning. Despite high token usage, when comparing token usage to cost, Kimi K2 Thinking becomes more competitive, costing around $380 for the benchmark run, comparable to Claude 4.5 Haiku, while outperforming it.

Coding and Writing Capabilities (7:12 - 11:06)

In coding tests, Kimi K2 showed mixed results. While it could generate files, it struggled with implementation details, such as rendering components on a Next.js page, leading to an unfunctional UI. The presenter notes it's not the best coding model for implementation and can be slow, with one request taking 10 minutes. Conversely, Kimi K2 excels in English writing quality. A test prompt asking for a defense of Java yielded a compelling and nuanced response, outperforming GPT-5 and Sonnet, which produced verbose, bullet-point heavy answers. This superior writing is attributed to Moonshot's focus on writing quality, even for an English-language model from a Chinese team.

Advanced Features and Open Model Landscape (11:06 - 15:54)

Kimi K2 is praised for its potential as a planning model, especially with tools like Kilo, allowing for multi-model task execution. It also performed well on the Skatebench benchmark, scoring 60% and becoming the best open-weight model for naming skate tricks, a task where many Chinese models typically struggle. The model supports "interled thinking for agentic tool use," a pattern where the model can reason during a reply without starting a new one, a feature also found in Claude and Miniax. This capability, along with its ability to execute hundreds of tool calls, is seen as a significant advancement for open models, putting pressure on American labs like OpenAI and Google.

Licensing and Future Outlook (15:54 - 23:57)

The video highlights a crucial aspect of Kimi K2: its modified MIT license. This license includes a clause requiring commercial products or services using Kimi K2 (or derivatives) to prominently display "Kimi K2" on their user interface if they exceed 100 million monthly active users or $20 million in monthly revenue. The presenter finds this reasonable, acknowledging that most services would likely display the model name anyway. The article also points out that open models are releasing faster, closing the performance gap with closed models. Chinese labs are noted for their rapid progress, with Kimi K2 being a prime example of their focus on benchmarks and user experience. The model's training cost was $4.6 million for one run, and its release as an open-weight model is seen as a significant advancement for the industry, despite initial challenges in hosting and vendor verification due to its complex tool-calling behaviors.

Conclusion (23:57 - 24:09)

The presenter expresses significant excitement for Kimi K2, planning to use it as a daily driver in T3 chat. They reiterate its status as a genuinely advanced, state-of-the-art open-weight model, despite some current limitations in code output, and encourages viewers to try it and share their opinions.

Notable Quotes

"This is the best tool calling model ever from what I have seen." — Theo

"The model's kind of nuts. is able to do 200 to 300 tool calls consecutively without a human interfering." — Theo

"The model's 594 gigs. It's a trillion parameters, which makes it the biggest openweight model ever made." — Theo

"Why is a Chinese model better at writing English in a compelling way than American-based companies making similar models?" — Theo

"If the software or any derivative works thereof is used for any of your commercial products or services that have more than a 100 million monthly active users or more than 20 million US or equivalent to other currencies in monthly revenue, then you shall prominently display Kimmy K2 on the user interface of such product or service." — Theo (quoting the license)

"Kimmy K2 thinking can execute up to two to 300 sequential tool calls without human interference, reasoning coherently across hundreds of steps to solve complex tasks." — Theo (quoting a blog post)