Kimi K3.1 Pro: Open Weights, Almost Opus
Read the Portuguese translation →
Author: Prompt Engineering · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Kimi K3.1 Pro: Open Weights, Almost Opus · Channel: Prompt Engineering · 01:02 · Watch on YouTube
Topics: AI evaluation and benchmarks, AI models, Local and open-source models, AI safety
Summary
Section titled “Summary”Moonshot released Kimi K3.1 Pro, an open-weights mixture-of-experts model with 2.8 trillion total parameters and about 104 billion active. The company says additional reinforcement learning brought its performance close to leading closed-source models, while its cost per task is about a third of Claude Opus 5.5’s. A shared demonstration showed the model inspecting a walking simulator and finding 36 errors in the shoulder muscle math. Moonshot removed cyber-related training data before release, but says the model still solves more than twice as many exploit tasks as Kimi K3; it is available on the Kimi API, with open weights due by November 4.
Key points
Section titled “Key points”- Kimi K3.1 Pro is an open-weights mixture-of-experts model with 2.8 trillion total parameters and about 104 billion active.
- It is the same size as Kimi K3, but was trained with substantially more reinforcement learning.
- Moonshot claims it matches or approaches leading closed-source models.
- On Terminal Event 4.0, it scores less than a point behind Claude Opus 5.5 and ahead of GPT-6 Astra.
- Moonshot’s cost chart puts its cost per task at about one-third of Opus’s.
- In a walking simulation test, it checked the simulator when progress stalled and reportedly found 36 errors in the shoulder muscle math.
- Moonshot removed cyber-related data from training before the open release because exploit skills kept increasing during training.
- Even after that change, Moonshot says it solves more than twice as many exploit tasks as Kimi K3.
- The model is live on the Kimi API at half Opus’s per-token price, and its open weights are expected by November 4.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Kimi K3.1 Pro and benchmark claims | The video introduces the model’s scale, reinforcement learning, and claimed performance against leading closed-source models. |
| 00:31 | Cost and simulator demonstration | Moonshot’s chart compares task costs, and a walking simulation example shows the model finding errors in the simulator. |
| 01:02 | Cybersecurity and availability | The video covers the model’s exploit-task performance, its API pricing, and the planned open-weights release. |
Quotes
Section titled “Quotes”It’s an open-source mixture of expert model with 2.8 trillion total parameters and about 104 billion active. — 00:00
it gets there for about a third of what Opus costs per task — 00:31
Moonshot says it found 36 errors in the shoulder muscle math. — 00:31
it still solves more than twice as many exploit tasks as Kimi K3 did. — 01:02
the open weights are coming by November 4. — 01:02
Concepts and entities
Section titled “Concepts and entities”Kimi K3.1 Pro, Moonshot, Claude Opus 5.5, GPT-6 Astra, Terminal Event 4.0, Kimi API
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 So, Moonshot just released Kimi K3.1 Pro, and here’s all you need to know. It’s an open-source mixture of expert model with 2.8 trillion total parameters and about 104 billion active. It’s the same size as Kimi K3, but Moonshot trained it with a lot more reinforcement learning. And they claim it now matches or comes close to the best closed-source models. For example, on Terminal Event 4.0, it lands less than a point behind Claude Opus 5.5 and ahead of GPT-6 Astra. And
00:31 on Moonshot’s own cost chart, it gets there for about a third of what Opus costs per task, which is kind of incredible. Here is an example of what it can do. Moonshot also shared a test where it built a walking simulation of the human body, and when the walk stopped improving, it went back and checked the simulator itself. Moonshot says it found 36 errors in the shoulder muscle math. Now, the most interesting part is about cybersecurity. Moonshot removed cyber-related data from training before the open release because its exploit
01:02 skills kept climbing as they trained it. But even after that, it still solves more than twice as many exploit tasks as Kimi K3 did. It’s live today on Kimi API at half the per token price of Opus, and the open weights are coming by November 4. For more quick updates, subscribe to the channel.