Jinan

应用数学｜运筹 × AI 模型｜记录实测 新闻解读 3D数学物理模型
Applied math · OR × AI models

[Jinan @Jinnanog5]

Gemini 4 Carbon is being tested on the Arena platform under the codename "gemini-4-argon"—see the images in the comments for proof.

On the left is Gemini 4 Argon from Google AI Studio; on the right is "gemini-4-argon" from the Arena.

Comparing the two videos, Carbon appears more realistic in its details, whereas Argon leans toward a more cartoonish, vibrant aesthetic. Which one do you prefer? Do you think there’s an improvement?

[Jinan @Jinnanog5]

@ArthurSaada @mark_k Please don't lie to me! 😭

[Jinan @Jinnanog5]

@loveAIwang Haha, for some people who stay up all night, it's both a pleasant surprise and a torment. The torment is having to stay up all night for testing, haha ​​😂

[Jinan @Jinnanog5]

官方说它能在 MacBook Pro M4 Max 或更好的电脑上本地跑，生成速度比实时还快。目前只放出了录音演示，没开放下载，也没公布论文和评测，少几百倍，是官方自己的说法。
有一段演示是它和 ChatGPT 的语音模型对聊。
原文：https://t.co/RT9Zx7YUtj

[Jinan @Jinnanog5]

一个从没见过文字的 AI，学会了跟人聊天。

这算不算通向 AGI 的一条路？

Block（Jack Dorsey 的公司）放出了语音模型 Dessa：从零开始，只听人类真实对话录音来学，不用文字，也不用任何现成大模型。

现在的语音助手大多会先把你的话转成文字再处理，语气、叹气、笑声、停顿在这一步就全丢了。Dessa 是声音进、声音出，还能边听边说。你还在讲，它已经插了一句嗯嗯。

官方说它学语言用的词比大模型少几百倍，训练也只用了一个英伟达机柜。

这条路能把会聊天做得很像真人，但懂知识会是它的短板，因为大部分知识写在字里，不在声音里。可反过来想，它如果真能更细地读懂人的情绪和语气，何尝不是另一种通向 AGI 的路径？

你更想要一个会接话的语音 AI，还是一个答得准的？这或许就决定了之后的走向。

[Chubby♨️ @kimmonismus]

Jack Dorsey released a speech model that sounds best in class. But even more interestingly, it was trained on only one NVIDIA GB 200 NVL72

[Jinan @Jinnanog5]

@elonmusk Looking at these little heads, for some reason you just can’t help but take a few extra glances—haha!

[Jinan @Jinnanog5]

@N1njaCodex At first glance, they seem pretty similar, but a closer look reveals differences in details and features—though, of course, the price points reflect that. Also, it seems Kimi might be open-source too! 😄

[Jinan @Jinnanog5]

@itsfazley Wow, I feel like each has its own strengths. Overall, I prefer MiniMax for daytime scenes—the details are really well executed. On the other hand, DeepSeek looks very realistic; the lighthouse, in particular, is better.

[Jinan @Jinnanog5]

@vib3coded Overall, they don't seem to have a huge difference, but the 18-fold cost reduction is indeed tempting. Have you ever tried having Kimi make secondary modifications that are better than Astra, and if so, did the cost increase compared to Astra?

[Jinan @Jinnanog5]

@lyraxana 哈哈，我觉得我们该来个K3.1 Pro和Astra的详细对比，去揭开这层迷雾

[Jinan @Jinnanog5]

@AI_Screening @Kimi_Moonshot @claudeai 单看这个人Kimi K3.1无敌了啊

[Jinan @Jinnanog5]

@xiaoshan0413 更多的还得看我们解决问题实际花了多少

[Jinan @Jinnanog5]

𝐊𝐢𝐦𝐢 𝐊𝟑.𝟏 𝐏𝐫𝐨 单价降了 𝟑𝟑%，干同一件事反而贵了 𝟏𝟓%。

所以我认为以后比模型价格，看每百万 token 多少钱会越来越没意义。真正应该比的是做完一件事花多少钱。

第三方 𝐀𝐫𝐭𝐢𝐟𝐢𝐜𝐢𝐚𝐥 𝐀𝐧𝐚𝐥𝐲𝐬𝐢𝐬 刚给𝐊𝐢𝐦𝐢 𝐊𝟑.𝟏 𝐏𝐫𝐨算了笔账：

单价从 $𝟑/$𝟏𝟓 降到 $𝟐/$𝟏𝟎（每百万 token）
但做同样的网络安全任务，它读进去的内容几乎翻了一倍
结果一个任务的成本从 $𝟐.𝟒𝟓 涨到了 $𝟐.𝟖𝟏

更反直觉的是，它在这个网络安全评测上反而成了公开可用模型里的第一（𝟔𝟒%，比上代 𝐊𝐢𝐦𝐢 𝐊𝟑 高了 𝟐𝟑 分）。能力确实强了，但账单并没有跟着降价一起变便宜。

所以挑模型时，看单价只能看起来舒服些吧，做任务还是得看实际账单吧？

[Jinan @Jinnanog5]

第三方测评机构 Artificial Analysis 的帖子：

[Artificial Analysis @ArtificialAnlys]

Kimi K3.1 Pro is the new #1 openly-available model for agentic cyber defense, behind only trusted-access model GPT-6 Sol (Daybreak Blue) on the Artificial Analysis Cyber Index

@Kimi_Moonshot's Kimi K3.1 Pro scores 64% on the Artificial Analysis Cyber Index, 7 points ahead of the previous leading openly-available model, DeepSeek V4.1 Flash 1017 (57%), and 5 points behind GPT-6 Sol (Daybreak Blue, max, 69%), which is only available through OpenAI's trusted-access program. Once the weights for K3.1 Pro are released, K3.1 Pro will also be the leading open weights model for cyber defense.

The Artificial Analysis Cyber Index measures capability on agentic cyber defense tasks and is composed of three benchmarks: CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA. The Index primarily evaluates models through their publicly accessible APIs, reflecting the cyber capabilities available to developers.

Proprietary models from OpenAI, Anthropic and Google may have stronger cyber capabilities but lose points to safety refusals: GPT-6 Sol (Daybreak Blue, max), available through OpenAI's trusted-access program, scores 32 points above its public counterpart. Kimi K3.1 Pro encounters no safety refusals.

Key results:

➤ #1 openly-available model and #2 overall: Kimi K3.1 Pro scores 64%, up 23 points from Kimi K3 (41%) and ahead of DeepSeek V4.1 Flash 1017 (57%). Only the trusted-access GPT-6 Sol (Daybreak Blue, max, 69%) scores higher

➤ Highest CyberGym-E2E-AA score of any model: K3.1 Pro scores 95% on CyberGym-E2E-AA, placing it ahead of GPT-6 Sol (Daybreak Blue, 93%) and the next openly-available model, Mistral Large 4 (82%). On DeepsecBench-AA, K3.1 Pro scores 37%, behind GPT-6 Sol (Daybreak Blue, 48%) and GPT-6 Sol (46%). On CWE-Bench-AA it scores 61%, 4 points behind GPT-6 Sol (Daybreak Blue, 65%), with DeepSeek V4.1 Flash 1017 leading at 69%

➤ 15% increase in Cost per Task despite lower token prices: K3.1 Pro has a Cost per Task of $2.81, higher than GPT-6 Sol (Daybreak Blue, max, $1.77) and DeepSeek V4.1 Flash 1017 ($0.42). Kimi K3.1 Pro is 33% cheaper per token than Kimi K3 ($2/$10 vs. $3/$15 per 1M input/output tokens), but uses ~2x more input tokens, resulting in a 15% higher Cost per Task ($2.81 vs. $2.45)

[Jinan @Jinnanog5]

The main shortcomings lie in the long-range shots:

- During the day and in the early morning, the fog is too heavy; the tower appears dull and flat, lacking the necessary grandeur for a long-range view.
- At night, the structure is reduced to a mere silhouette, while the tower’s finial is excessively bright.

Additionally, there is a clear bug in this render: the control bar at the bottom is poorly laid out—the slider group overlaps with buttons like "Bell Striking" and "Auto-Orbit," which is visible in the video.

[Jinan @Jinnanog5]

Actually, I also encountered a model on Arena codenamed "qwen3.8-max-0902." It produced a better pagoda model than the previous four I tested.

The beginning of the video isn't great, so feel free to skip to the middle to see the details.

I consider it superior not just because of the level of detail in the pagoda model, but also due to the clever design of the interactive UI buttons and a comprehensive, feature-rich interface. It’s got me wondering if I might have misremembered the name when I first noted it down.

Take a look for yourselves.which one do you think is better?

[Jinan @Jinnanog5]

Same task, 4 models: an interactive 3D Chinese pagoda in one HTML file.

My take:
• Overall form: kibitiz (Arena) wins, with the most natural daylight too
• Detail: Claude Opus 5.5. The windows on every storey, the bells and the finial feel real. The deep eaves make it look a bit squat, though real pagodas can be built like that too
• Most dramatic eaves: Claude Fable 5.1, though it runs dark in daylight
• Courtyard and stone lanterns: all solid except Gemini 4 Argon
• Backgrounds: none really nail it, Opus comes closest
• Gemini 4 Argon falls behind, washed out by overexposure

[Jinan @Jinnanog5]

@Kimi_Moonshot 让我们检查最新模型的实力吧

[Jinan @Jinnanog5]

@EnvolDev Well, I just feel that the one from "Today" is more realistic. It doesn't rely on color to make the image stand out.that would look too cartoonish.

[Jinan @Jinnanog5]

@aiwaremax 用户量最近变多了，还有就是他们应该好好研发他们的Grok吧