Moonshot AI Debuts 3T-Parameter Kimi K3.1 to Challenge Claude Opus

Fireship

Moonshot AI's benchmark parity with Claude challenges Western assertions that distillation defenses halted Chinese frontier progress, even as the startup deliberately weakens its model's hacking capabilities to evade US sanctions.

AI-assisted summary of This 3 trillion parameter model should be impossible... — original source. Check the original for full context.

Source published: 2026-10-26

Key ideas

Moonshot AI announced Kimi K3.1, a massive 3-trillion-parameter open-weight model scheduled for full download on November 4th after secretly debuting under the alias 'Kibitz' on Chatbot Arena. (0:00)
Kimi K3.1 rivaled proprietary frontier models on coding evaluations, landing within a single point of Claude Opus 5.5 on TerminalBench 4.0 and claiming first place on SWE-marathon. (2:08)
To evade severe geopolitical blacklists and regulatory bans, Moonshot deliberately pruned offensive cyber exploit data, reducing Kimi's internal exploit benchmark pass rate from 90.5% to 81%. (2:08)
Practical procedural testing on Three.js 3D rendering revealed underlying limits: while Kimi produced photorealistic static meshes, it relied on crude mesh warping rather than skeletal inverse kinematics. (3:46)

Detailed notes

## Background Published by the tech channel Fireship on YouTube on October 26, 2026, the video investigates Moonshot AI's newly announced Kimi K3.1, a free 3-trillion-parameter open-weight model. The description highlights how the model came within one point of Claude Opus on Terminal Bench and questions whether its real-world coding and performance live up to the benchmark hype, alongside a promotion for Mobbin Pro.

## TL;DR The video breaks down the sudden emergence of Moonshot AI's Kimi K3.1—initially surfacing anonymously on Chatbot Arena as 'Kibitz'—which challenges Western claims that distillation defenses had crippled Chinese frontier AI development. While Kimi achieved elite scores on TerminalBench and SWE-marathon, Moonshot deliberately crippled its offensive cybersecurity exploits to avoid severe geopolitical sanctions and federal bans. Putting corporate benchmarks aside, Fireship evaluates the model using his custom 'Thoroughbench' Three.js 3D horse challenge, discovering that despite impressive visual aesthetics, Kimi relies on crude mesh warping rather than skeletal articulation and scores a modest 4.20. The episode wraps up by demonstrating how developers can use Mobbin's Model Context Protocol server to feed AI coding agents authentic design patterns instead of generic UI layouts.

## Atmosphere & style The presentation is punchy, irreverent, and fast-paced, marked by Fireship's signature deadpan humor and satirical tech commentary. It blends cynical analysis of international AI intellectual property disputes with playful, hands-on absurdity during the Three.js horse breakdown. The viewing experience feels like a sharp, insider developer briefing that refuses to take benchmark marketing at face value.

## Bulletin - [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) Moonshot AI announced Kimi K3.1, a massive 3-trillion-parameter open-weight model scheduled for full download on November 4th after secretly debuting under the alias 'Kibitz' on Chatbot Arena. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) Kimi K3.1 rivaled proprietary frontier models on coding evaluations, landing within a single point of Claude Opus 5.5 on TerminalBench 4.0 and claiming first place on SWE-marathon. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) To evade severe geopolitical blacklists and regulatory bans, Moonshot deliberately pruned offensive cyber exploit data, reducing Kimi's internal exploit benchmark pass rate from 90.5% to 81%. - [03:46](https://www.youtube.com/watch?v=G_em1hHPnXY&t=226s) Practical procedural testing on Three.js 3D rendering revealed underlying limits: while Kimi produced photorealistic static meshes, it relied on crude mesh warping rather than skeletal inverse kinematics.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) Moonshot AI secretly evaluated Kimi K3.1 on Chatbot Arena under the moniker 'Kibitz', a Yiddish term for onlookers offering unhelpful chess advice. - [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) Kimi K3.1 is scheduled for open-weight public release on November 4th as a 3-trillion-parameter architecture. - [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) Western frontier labs like Anthropic have accused Chinese firms of mass distillation attacks using swarms of automated accounts to harvest reasoning tokens. - [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) Fireship defends open distillation, arguing that frontier labs originally vacuumed up human intellectual property from blogs, code repositories, and physical books. - [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) Anthropic claimed to block 90% to 99% of distillation scraping from China, which contradicts assumptions that Chinese frontier parity relies entirely on stolen outputs. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) On TerminalBench 4.0, Kimi K3.1 came within one point of Claude Opus 5.5, with Artificial Analysis recording an independent score of 67%. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) Kimi K3.1 took top honors on SWE-marathon, an evaluation measuring performance on multi-hour enterprise software engineering tasks. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) Moonshot purged vulnerability data from pre-training, dropping offensive exploit capability from an unconstrained 90.5% down to 81% for public release. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) Tactical cybersecurity lobotomies are driven by geopolitical survival to prevent total bans across Western markets like the federal ban enacted on DeepSeek. - [02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) Kimi K3.1 topped Artificial Analysis's Cyber Index largely by default, as Claude and GPT-6 triggered safety refusals across all benchmark prompts. - [03:46](https://www.youtube.com/watch?v=G_em1hHPnXY&t=226s) In Fireship's custom 'Thoroughbench' Three.js 3D horse challenge, Kimi K3.1 scored 4.20/10, falling behind Claude Opus (6.39) and GPT-6 Astra (5.21). - [03:46](https://www.youtube.com/watch?v=G_em1hHPnXY&t=226s) While Kimi K3.1 excelled at static visual texturing and aesthetic anatomy, it animated meshes via crude distortion rather than true skeletal rigs. - [03:46](https://www.youtube.com/watch?v=G_em1hHPnXY&t=226s) Claude Opus demonstrated superior mechanical procedural generation by articulating 95,000 vertices across 35 independently animated bones. - [05:29](https://www.youtube.com/watch?v=G_em1hHPnXY&t=329s) Mobbin's Model Context Protocol (MCP) server integrates into tools like Cursor and Claude Code, providing access to 600,000+ real-world app screens to avoid generic AI interface designs.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) The "Kibitz" Mystery and the Distillation Paradox

The episode begins with the sudden appearance of an anonymous frontier model codenamed "Kibitz" on Chatbot Arena. The name itself is borrowed from the Yiddish term for a spectator leaning over a chess player's shoulder offering bad advice on how to blunder a queen—an analogy Fireship calls the quintessential description of an AI coding assistant. The mystery did not last long, as the opening letters gave away the true creator: Moonshot AI. Right on cue, Moonshot officially announced Kimi K3.1, a massive 3-trillion-parameter open-weight model scheduled for full download on November 4th. The announcement dropped right when industry observers assumed the generative AI hype cycle was deflating and benchmark inflation had reached its ceiling.

Fireship frames the release against the backdrop of bitter intellectual property disputes between Western AI titans and Chinese labs. Earlier in the year, Anthropic accused Moonshot AI, alongside firms like DeepSeek and MiniMax, of executing coordinated distillation attacks against Claude. These efforts allegedly involved sending swarms of automated accounts to harvest reasoning traces and raw output tokens, which were subsequently used to train free open-weight models. While OpenAI and Anthropic view unauthorized distillation as "grand theft autocomplete" and have poured millions into mitigation, Fireship defends model harvesting on ideological grounds: “And now, I believe distillation is a fundamental human right because big AI companies have literally distilled a collective human brain.” Frontier labs trained their base models by vacuuming up public web blogs, GitHub repositories, academic literature, and even destructively scanning physical books via initiatives like Project Panama.

The paradox centers on Anthropic’s recent claims that automated distillation has effectively been solved, with 90% to 99% of scraping attempts from China successfully blocked. If frontier model makers have successfully shut down data harvesting at both massive scales and small-scale operations, a Chinese open-weight model performing at Kimi's level should theoretically be impossible under Western assumptions. For years, established labs claimed Chinese engineering parity was merely a byproduct of stolen tokens; if that supply chain is severed, Kimi K3.1’s sudden emergence at the top of the leaderboards challenges the narrative entirely.

```mermaid flowchart TD accTitle: The Distillation Dispute Flow accDescr: Diagram showing how frontier labs harvest public human data and how open-weight labs distill frontier outputs.

subgraph DataCollection["1. Public Data Harvesting"] A["Collective Human Brain<br/>(Web blogs, GitHub, scanned books)"] B["Frontier Closed Labs<br/>(OpenAI, Anthropic)"] A -->|"Destructive scanning & web scraping"| B end

subgraph DistillationDispute["2. Distillation & Defense"] C["Open-Weight Competitors<br/>(Moonshot AI, DeepSeek, MiniMax)"] B -->|"Reasoning tokens & output swarms"| C B -.->|"90–99% blocked via distillation defenses"| C end

subgraph ModelRelease["3. Final Model"] D["Kimi K3.1<br/>(3T Parameter Open-Weight Model)"] C -->|"Trained & released"| D end ```

- **Anonymous testing moniker:** The model debuted on Chatbot Arena under the name "Kibitz." - **Model specifications:** Kimi K3.1 is reported as a 3-trillion-parameter open-weight model developed by Moonshot AI, slated for full download on November 4th. - **Distillation defense claims:** Anthropic reported neutralizing between 90% and 99% of inbound distillation requests originating from China.

---

[02:08](https://www.youtube.com/watch?v=G_em1hHPnXY&t=128s) Benchmarks and Deliberate Cyber-Lobotomy

On raw metrics, Kimi K3.1 posts numbers that place it in direct competition with closed proprietary systems. On TerminalBench 4.0, the model landed within a single point of Claude Opus 5.5—a threshold unmatched by any competing open-weight model. Independent replication by Artificial Analysis confirmed Kimi scored 67%, the highest mark recorded across their suite. However, testing harnesses remain a critical variable that can dramatically alter evaluations across different testbeds. The model also claimed top placement on SWE-marathon, an evaluation benchmark designed around multi-hour, enterprise-grade engineering specifications that simulate tasks requiring weeks of human labor without AI assistance.

Despite these coding achievements, Moonshot made a deliberate choice to cripple the system's offensive cybersecurity utility. The development team actively purged security exploits and vulnerability data from the pre-training corpus to avoid producing an AI capable of automated hacking. Explaining this tactical self-censorship, Fireship observes: “And that's why Moonshot really has no choice but to lobotomize its hacking capabilities.” Internal testing showed that while an unconstrained checkpoint completed 90.5% of illegal exploit challenges, the sanitized public release dropped to 81%.

This sanitization serves as a geopolitical survival tactic. Following the United States government’s decision to ban DeepSeek across all federal civilian agencies, Chinese labs face the perpetual threat of comprehensive blacklists that would grant domestic incumbents like OpenAI and Anthropic an uncontested market advantage. Ironically, even with its defenses artificially degraded, Kimi K3.1 still ranked first on Artificial Analysis's Cyber Index, largely because Western rivals like Claude and GPT-6 simply triggered automated safety refusals on every prompt and defaulted to zero.

- **TerminalBench 4.0:** Kimi K3.1 finished within one point of Opus 5.5, with Artificial Analysis recording an independent score of 67%. - **SWE-marathon:** Kimi secured first place on long-horizon, multi-hour engineering tasks. - **Cybersecurity filtering:** Internal unfiltered checkpoints achieved 90.5% on exploit benchmarks, which dropped to 81% after Moonshot stripped sensitive data to avoid international regulatory backlash. - **Cyber Index standing:** Kimi captured the top position by default after Claude and GPT-6 triggered safety refusals across all benchmark prompts.

---

[03:46](https://www.youtube.com/watch?v=G_em1hHPnXY&t=226s) Thoroughbench: Testing Kimi K3.1 with Three.js 3D Horses

Setting aside corporate benchmarks, Fireship subjects Kimi K3.1 to his personal, state-of-the-art evaluation metric: "Thoroughbench." Honed over thousands of hours of research, the benchmark evaluates an AI model on a single prompt entered directly into the chat interface: `create a realistic interactive 3D horse with Three.js`. The output is scored from 0 to 10 against a proprietary mental loss function. Previous frontier models set the baseline: Google Argon posted an abysmal 0.69, GPT-6 Astra scored a 5.21, and Claude Opus took the lead with a 6.39.

After finding the free tier blocked, Fireship shelled out 250 yuan per month for the premium subscription to run the test on Kimi K3.1. Visually, the model generated what appeared to be the most aesthetic horse of the entire group. Its texturing and colors were realistic, and the anatomical proportions—down to its copulatory organ—were rendered tastefully. However, initiating the walking animation exposed fundamental architectural shortcomings. Emphasizing the mechanical flaws under the hood, Fireship notes: “If we look closely, you can see it achieves the animation by simply warping the original mesh, which is a very unsophisticated and half-assed way to animate a 3D object.”

In contrast, Claude Opus constructed an intricate mesh featuring 95,000 individual vertices bound to an internal skeleton of 35 distinct, independently articulated bones. While Gemini Argon bizarrely included a redundant bone in its rig, both Western models relied on true skeletal rigging to avoid structural collapse during movement. Because Kimi relied on brute-force mesh warping instead of inverse kinematics, its animation appeared as if the horse were being sucked into a gravitational singularity. Kimi K3.1 earned a final Thoroughbench score of 4.20—respectable, but still behind the reigning frontier giants.

---

[05:29](https://www.youtube.com/watch?v=G_em1hHPnXY&t=329s) Agent UI Design with Mobbin MCP

Transitioning from 3D procedural modeling to front-end development, Fireship shifts the discussion to UI design quality and introduces episode sponsor Mobbin. While coding agents can spin up boilerplate layouts, prompting standard tools for a login screen typically produces generic "purple gradient vibe slop." Mobbin addresses this creative stagnation through its Model Context Protocol (MCP) server, which bridges coding agents directly into an index containing more than 600,000 production interface screens gathered over years of real-world app releases.

By plugging the MCP server into agents like Claude Code, Cursor, or Codex, developers can direct their models to analyze real workflow patterns before generating code. The agent inspects how top-tier commercial applications handle end-to-end user flows, generates a synthesized visual report of shared interface paradigms, and adapts the architecture to the developer's project. Highlighted features include:

- **600,000+ UI screens:** An extensive reference library of production interfaces used to ground autonomous agents in proven UX patterns. - **Traceable lineage:** Generated components link directly back to their source application screens for manual verification and design auditing. - **Agent compatibility:** Integrates directly via MCP into developer tools such as Cursor, Claude Code, and Codex. - **Promotional offer:** A 20% discount on Mobbin Pro provided via the sponsor link in the description.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=G_em1hHPnXY&t=0s) I believe distillation is a fundamental human right because big AI companies have literally distilled a collective human brain. — Fireship

## Mentioned Moonshot AI, Kimi K3.1, Anthropic, OpenAI, Claude, DeepSeek, MiniMax, Chatbot Arena, Project Panama, TerminalBench 4.0, Artificial Analysis, SWE-marathon, GPT-6, Three.js, Google Gemini Argon, GPT-6 Astra, Mobbin, Claude Code, Cursor, Codex

## Video description Get 20% off Mobbin Pro to help your agent design UIs that actually convert - https://mobbin.com/fireship

Kimi K3.1 just dropped as a free 3-trillion-parameter open-weight model that comes within one point of Opus on Terminal Bench. But is it *actually* good? Let's find out.

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=G_em1hHPnXY)

中文简报

月之暗面推出3万亿参数Kimi K3.1挑战Claude Opus 月之暗面在基准测试中与Claude旗鼓相当，挑战了西方关于蒸馏防御已遏制中国前沿进展的论断，尽管这家初创公司为规避美国制裁故意削弱了该模型的黑客攻击能力。本视频剖析了月之暗面（Moonshot AI）Kimi K3.1的突然崛起——该模型最初以“Kibitz”为代号匿名亮相Chatbot Arena——挑战了西方认为蒸馏防御已阻碍中国前沿AI发展的说法。尽管Kimi在TerminalBench和SWE-marathon上取得了顶尖成绩，但月之暗面故意削弱了其进攻性网络安全利用能力，以规避严厉的地缘政治制裁和联邦禁令。抛开企业基准测试，Fireship使用其自创的“Thoroughbench”Three.js 3D马匹挑战对其进行了评估，发现尽管视觉美感令人印象深刻，但Kimi依赖于粗糙的网格变形而非骨骼关节点动画，仅获得了4.20的普通评分。本期视频最后展示了开发者如何利用Mobbin的模型上下文协议（MCP）服务器为AI编程智能体提供真实的优秀设计模式，而非千篇一律的通用UI布局。月之暗面宣布推出Kimi K3.1，这是一款拥有3万亿参数的巨型开放权重模型，在以别名“Kibitz”暗中亮相Chatbot Arena后，计划于11月4日提供完整下载。 (0:00) Kimi K3.1在代码评测中与顶尖闭源前沿模型相匹敌，在TerminalBench 4.0上与Claude Opus 5.5仅差1分，并在SWE-marathon上夺得第一名。 (2:08) 为了规避严厉的地缘政治黑名单和监管禁令，月之暗面故意删减了攻击性网络漏洞利用数据，将Kimi的内部漏洞利用基准通过率从90.5%降至81%。 (2:08) 在Three.js 3D渲染的实际程序化测试中显露了底层局限：虽然Kimi生成了逼真的静态网格，但它依赖粗糙的网格变形而非骨骼反向动力学。 (3:46)

Reading summary

Moonshot AI's release of the 3-trillion-parameter open-weight model Kimi K3.1 challenges Western assumptions that Chinese AI progress relies strictly on unauthorized distillation from frontier labs. While the model rivals proprietary systems in software engineering, technical stress tests and deliberate capability pruning reveal ongoing limitations and geopolitical constraints. The Distillation Paradox Behind Kimi K3.1 (0:00) After initially appearing anonymously under the moniker "Kibitz" on Chatbot Arena, Moonshot AI announced Kimi K3.1, a massive 3-trillion-parameter open-weight model scheduled for release on November 4th. The announcement surfaced amid intense intellectual property disputes, including Anthropic's accusations that Chinese firms like Moonshot, DeepSeek, and MiniMax deployed automated swarms to distill Claude's outputs. Fireship contrasts corporate theft claims with the reality that frontier labs built their own foundational models by scraping public web text, code repositories, and scanned books. Anthropic recently claimed to have neutralized 90% to 99% of inbound distillation traffic originating from China. Fireship argues that if Western labs have effectively closed data harvesting pipelines, Kimi K3.1’s emergence near the top of public leaderboards undermines the narrative that Chinese AI performance is merely a derivative product of stolen tokens. Coding Parity and Deliberate Cyber Sanitization (2:08) In benchmark testing, Kimi K3.1 demonstrated coding capabilities comparable to closed Western frontier models. It landed within a single point of Claude Opus 5.5 on TerminalBench 4.0—validated independently by Artificial Analysis with a 67% score—and captured first place on SWE-marathon's multi-hour engineering evaluations. Despite its technical competence, Moonshot intentionally removed offensive cybersecurity data from the pre-training set. Unfiltered checkpoints completed 90.5% of exploit challenges, but Moonshot reduced this capability to 81% in the public model. Fireship attributes this self-imposed constraint to geopolitical survival: following the U.S. federal ban on DeepSeek, Moonshot sought to avoid international blacklists. Ironically, Kimi K3.1 still ranked first on Artificial Analysis's Cyber Index because rivals such as Claude and GPT-6 triggered safety refusals and scored zero. Architectural Deficits in Procedural 3D Generation (3:46) Fireship subjected Kimi K3.1 to an informal procedural test, dubbed "Thoroughbench," which evaluates how models code an interactive 3D horse using Three.js from a single chat prompt. Although Kimi produced visually appealing textures and anatomically proportioned static renders, its animation mechanics revealed significant technical limitations compared to Western competitors. While Claude Opus constructed a complex mesh of 95,000 vertices mapped to a 35-bone internal skeletal rig, Kimi K3.1 lacked skeletal rigging and attempted movement by warping the outer mesh. This brute-force approach resulted in visual distortion during animation, earning Kimi a score of 4.20 out of 10—behind GPT-6 Astra (5.21) and Claude Opus (6.39). Grounding Coding Agents in Production UI (5:29) To address repetitive visual outputs in automated front-end development, Fireship highlights the Model Context Protocol (MCP) server from episode sponsor Mobbin. While generalist coding models frequently default to generic interface templates, the MCP server connects agents directly to an indexed repository of over 600,000 production interface screens. Tools such as Cursor, Codex, and Claude Code can query this index to review authentic enterprise design patterns and end-to-end user flows. By inspecting real-world implementations, agents generate interface components that retain direct lineage to original application references rather than improvising ungrounded layouts. Kimi K3.1 demonstrates that open-weight models can achieve near-parity with top proprietary systems in standardized software engineering tasks, even amid defensive data filtering by Western labs. However, architectural shortcuts in complex rendering tasks and deliberate feature pruning highlight how technical and political pressures continue to shape model performance.

精炼阅读摘要

Moonshot AI 推出拥有 3 万亿参数的开源权重模型 Kimi K3.1，在多项前沿编码与长周期工程基准测试中展现出与西方顶尖闭源模型比肩的性能；Fireship 指出，该模型的突破不仅对西方实验室关于“蒸馏封锁已切断技术追赶”的论调提出了核心质疑，也反映出开源模型在底层工程架构与地缘合规约束下的多重博弈。知识蒸馏争议与地缘技术追赶悖论 (0:00) Kimi K3.1 最初以代号“Kibitz”在 Chatbot Arena 匿名亮相，随后 Moonshot AI 正式公布了这款参数量达 3 万亿的开源权重模型。此前，Anthropic 等西方前沿实验室指责 Moonshot AI、DeepSeek 和 MiniMax 等中国机构利用自动化账号集群抓取 Claude 的推理链路与原始标记，将其定性为未授权的蒸馏盗用，并声称已建立防线成功拦截了 90% 至 99% 来自中国的抓取请求。然而 Fireship 认为，闭源实验室自身早期亦是无偿吸纳全网博客、GitHub 仓库及实体书籍数字化数据来构建基础模型，蒸馏本质上是技术获取的正当路径。这一交锋构成了核心的技术悖论：如果西方实验室关于蒸馏防御已基本奏效、流量通道被切断的断言成立，那么在西方既定认知中，完全依赖“数据窃取”实现同级别性能的中国开源模型在理论上本不可能出现。Kimi K3.1 的突围表明，要么封锁防御未能如西方所称般全面生效，要么开源实验室已建立起不单纯依赖西方模型输出的独立工程体系，直接挑战了既有的地缘技术叙事。工程基准突破与网络能力的主动阉割 (2:08) 在标准化评测中，Kimi K3.1 展现出比肩前沿闭源系统的编码能力。在 TerminalBench 4.0 测试中，其表现与 Claude Opus 5.5 仅差 1 分，并在独立机构 Artificial Analysis 的复现测试中取得该评测体系下最高的 67%；在模拟长周期、多小时企业级工程任务的 SWE-marathon 上，该模型同样斩获榜首。这些指标表明其在复杂逻辑执行与持久软件工程工作流方面已达到工业化竞争水准。与此同时，Moonshot AI 在网络安全攻防能力上做出了显著的战术性自抑。为规避海外严格的监管审查和潜在制裁风险，研发团队在预训练阶段主动剔除了系统利用漏洞与攻击性数据，导致其在漏洞挑战中的完成率从未过滤内部版本的 90.5% 下滑至公开发行版的 81%。尽管网络攻击能力被刻意弱化，但由于 Claude 和 GPT-6 等西方模型在相关基准测试中因安全拦截机制全面触发拒答并计为零分，Kimi K3.1 最终反而在 Artificial Analysis 网络安全指数榜单中位列第一。程序化图形生成暴露底层力学短板 (3:46) 脱离标准化基准测试，Fireship 采用了其自设的“Thoroughbench”评测指标——通过单一提示词要求模型使用 Three.js 生成具备交互功能的逼真 3D 跑马，以检验其底层数学建模与物理运动学逻辑。在这一测试中，Claude Opus 凭借 95,000 个顶点及由 35 根独立骨骼构成的反向运动学骨架获得 6.39 分，GPT-6 Astra 获得 5.21 分，而 Google Gemini Argon 因骨骼绑定冗余仅获 0.69 分。Kimi K3.1 最终获得 4.20 分。测试结果显示，Kimi K3.1 在静态渲染和色彩比例上呈现出极佳的视觉真实感，但其在动态交互中暴露出明显的机械结构缺陷。该模型并未构建真正的内部骨骼绑定系统，而是依赖粗暴扭曲原有网格（Mesh Warping）的方式来实现行走动画，导致整体动态表现极不协调。这一局限揭示出该模型尽管具备庞大的参数量与表面学习能力，但在需要精确几何物理理解与复杂分层结构控制的场景下，与西方头部闭源架构依然存在代际差距。上下文协议与界面设计的工程演进 (5:29) 针对当前 AI 编程智能体在前端用户界面开发中容易生成千篇一律的同质化模板问题，技术生态正通过模型上下文协议（Model Context Protocol，简称 MCP）探索外部架构扩展。以集成逾 60 万张真实商业级应用界面的 Mobbin MCP 服务器为例，该方案能够将外部成熟设计库直连 Cursor、Claude Code 与 Codex 等智能体工具，协助模型在生成代码前先消化经过生产检验的交互范式。这种架构强化模式使开发者能够引导智能体拆解端到端的用户交互路径，不仅生成具备系统性设计规范的前端组件，还能实现溯源审计以校验设计决策的真实出处。这种从静态参数生成向外挂结构化生产经验调用的演进，为弥补基础模型在专业细分领域审美或工程结构感知不足提供了一条现实的技术路径。 Kimi K3.1 的发布证伪了单凭数据封锁即可阻绝开源大模型追赶的假设，其基准成绩彰显了开源阵营在通用编码领域的跨越；然而，3D 程序化生成中显露的底层运动学缺陷与为了应对监管而采取的主动安全阉割，证明庞大的参数体量仍难以完全掩盖核心算法架构的短板与地缘政治约束。

Source quotes

[0:00] Fireship: I believe distillation is a fundamental human right because big AI companies have literally distilled a collective human brain.

Original transcript

[0:00] Last week, an anonymous mystery model [0:02] named Kibitz showed up randomly on [0:04] Chapot Arena performing at the frontier [0:06] level. Kibitz is the Yiddish word for [0:08] that guy standing behind you at a chess [0:10] game, it giving you tips on how to [0:11] blunder your queen, which to be honest [0:13] is the perfect analogy for a coding [0:14] agent. But it was pretty obvious who was [0:16] behind this mystery model from the [0:18] beginning because the first two letters [0:19] of this word are a dead giveaway. [0:21] Everybody knew it was Kimmy. Then right [0:23] on Q last Friday, Moonshot dropped Kimmy [0:25] K3.1, officially giving us a new three [0:28] trillion parameter openweight model that [0:30] will be fully downloadable on November [0:32] 4th. Or so they say. And just when you [0:34] thought the AI bubble was popping and [0:36] the Trust Me Bro benchmarks could not [0:37] possibly get any better, Kimmy made them [0:39] better. [0:40] >> I can go lower. [0:41] >> In today's video, we'll find out why [0:42] Kimmy K3.1's performance should be [0:44] impossible and put it to the test on the [0:46] only benchmark that matters, the [0:48] building 3D horses with 3JS. It is [0:50] October 26, 2026, and you're watching [0:53] the Code Report. Earlier this year, [0:55] Anthropic accused Moonshot AI, the [0:57] company behind Kimmy, alongside a bunch [0:59] of other Chinese companies like Deep [1:01] Seek and Miniax of running large-scale [1:03] distillation attacks against Claude by [1:05] sending swarms of fake users to steal [1:07] Claude's output and reasoning tokens, [1:09] which could then be used to train these [1:10] open models that they give to us for [1:12] free. And now, I believe distillation is [1:14] a fundamental human right because big AI [1:16] companies have literally distilled a [1:17] collective human brain. but by taking [1:19] our internet web blogs, our GitHubs, our [1:21] scientific papers, and they even [1:23] destructively scanned our books without [1:25] asking. Just Google Project Panama. But [1:27] OpenAI and Anthropic don't really see [1:29] distillation as a human right like I do. [1:31] To them, it's grand theft autocomplete. [1:33] And they've invested millions and [1:34] millions of dollars into stopping it. [1:36] And just a few weeks ago, Anthropic said [1:38] it pretty much solved the distillation [1:40] problem with 90 to 99% of distillation [1:42] attacks from China now being blocked. [1:44] And over at OpenAI, they've proved that [1:47] their tech can even stop small-scale [1:48] distilleries like PewDiePie. But if [1:50] that's true, that means an open Chinese [1:52] model with Kimmy's performance should [1:54] not be possible. Because for years, Big [1:56] AI has insinuated that the reason [1:58] Chinese models get so good so fast is [2:00] that they steal Big AI's outputs. But if [2:03] we can trust this Trust Me Bro [2:04] distillation data from Anthropic, that [2:06] would mean that these Trust Me Bro [2:07] benchmarks from Kimmy should not be [2:09] possible. like Kimmy K 3.1 is within one [2:12] point of opus 5.5 on terminal bench 4.0. [2:15] That's pretty crazy. And no other open [2:17] model is remotely close. And we might [2:19] even be able to actually trust these [2:21] numbers because artificial analysis ran [2:23] it independently and got 67% which is [2:26] the highest score of any model they [2:27] tested. But one thing to keep in mind is [2:29] that the harness plays a big role here. [2:31] Running different models with different [2:32] harnesses will produce huge differences [2:34] in benchmark results. In addition, Kimmy [2:37] also wins on S. marathon, which is a [2:39] test for multi-our software [2:41] requirements, or in other words, jobs [2:43] that would take humans weeks of work to [2:45] finish without AI assistance. That's all [2:47] pretty impressive, but there is one big [2:48] caveat when it comes to coding and [2:50] specifically cyber security. Moonshot [2:53] deliberately ripped out a bunch of cyber [2:54] security data from its training data [2:56] set, where the goal was to make the [2:58] model even worse at hacking than pirate [2:59] software. [3:00] >> I'm a genius, dude. And it worked [3:02] because internally at Moonshot, an [3:04] unfiltered checkpoint solved 90.5% of [3:07] their illegal exploit tests. But the [3:09] final release model only solves 81%. And [3:12] that's a smart business move on their [3:13] part because just a few days ago, the US [3:15] banned China's Deepseek model from [3:17] government use, including all federal [3:19] civilian agencies. And they would just [3:21] love a good reason to ban all Chinese [3:22] models across the board to hand Open AAI [3:25] and Anthropic a huge win and make all [3:27] the insiders even more unbelievably rich [3:29] than they already are. And that's why [3:30] Moonshot really has no choice but to [3:32] lobomtomize its hacking capabilities. [3:34] But despite that, Kimmy still performed [3:36] on top of all publicly available models [3:38] on the new cyber index at artificial [3:40] analysis. Although that's mostly because [3:42] Claude and GBT6 declined every prompt on [3:44] safety grounds and just took the L. But [3:46] nowadays, the only AI capability that I [3:48] really care about is how well it can [3:50] generate a 3D horse with 3JS. And to [3:53] analyze the skill objectively and [3:54] scientifically, I use a formal [3:56] benchmark. It's a state-of-the-art eval [3:58] I've developed after thousands of hours [3:59] of research, which I call thoroughbench. [4:02] The way it works is I literally just go [4:03] into the chat window and type create a [4:05] realistic interactive 3D horse with 3JS [4:08] and then press enter. From there, I rank [4:10] them from 0 to 10 based on a proprietary [4:12] loss function I keep in my brain. Using [4:14] this technique, Google Argon achieved a [4:17] 0.69. GPT6 Astra achieved a 5.21 and [4:22] Opus achieved a 6.39. But now it's time [4:25] to try Kimmy K3.1. I tried to be cheap [4:27] and use the free tier, but got denied. [4:29] So now I'm paying 250 yuan per month to [4:31] show you guys this demo. It was worth it [4:33] though because Kimmy produced what is [4:35] easily the most beautiful horse in this [4:36] sample. Its coloring is spot-on. It [4:38] looks realistic. And most importantly, [4:40] its copulatory organ is a reasonable [4:42] size. I was blown away by its beauty at [4:44] first glance, but then I pushed play on [4:46] the animation and found a problem. The [4:49] animations are absolutely terrible. If [4:51] we look closely, you can see it achieves [4:52] the animation by simply warping the [4:54] original mesh, which is a very [4:56] unsophisticated and half-assed way to [4:58] animate a 3D object. If we compare that [5:00] to Claude, the horse itself may not look [5:02] as beautiful, but it actually built a [5:04] mesh with 95,000 vertices with each [5:07] shape tied to one of 35 bones, and all [5:09] those bones are animated independently. [5:11] Even Gemini Argon used bones, but for [5:14] some reason, it added an extra one. The [5:15] skeletal animation is more difficult, [5:17] but it's why the horse doesn't look like [5:19] it's being torn apart by a black hole [5:20] when it walks. So, the final score for [5:22] Kimmy K 3.1 on Thoroughbench is a 4.20, [5:26] which is pretty dang impressive, but [5:28] still behind the other Frontier models. [5:29] But if you really want to build an [5:31] amazing 3D horse or user interface, you [5:33] need to check out Mobin, the sponsor of [5:35] today's video. I've been using Mobin for [5:37] over 5 years and its MCP server now lets [5:40] your coding agent study its library of [5:42] 600,000 real UI screens before it [5:45] designs anything. So instead of asking [5:47] for a login page and getting the same [5:49] purple gradient vibe slop, you start by [5:51] telling Claude Code, Cursor, or Codeex [5:53] to study how the best apps built that [5:55] complete user flow. It'll use Mobin to [5:57] gather references and build a visual [5:59] report of the patterns they all share. [6:01] Then it'll recreate that flow and adapt [6:04] it to your app. Every result points back [6:06] to the original screen. This so you can [6:08] check the source and after you connect [6:09] Mobin, everything your agent designs [6:11] will feel like it has actual taste. But [6:13] try Mobin for free today and get a 20% [6:16] discount at the link below. This has [6:18] been the code report. Thanks for [6:19] watching and I will see you in the next [6:21] one.