Mistral Tops Reflection AI's Beam to Lead Global Open-Weight AI Race

Fireship

Surveillance and espionage risks tied to proprietary cloud APIs are forcing governments and enterprises toward open-weight models despite unreleased weights and disputed benchmarks.

AI-assisted summary of A $6.3 billion open-weight model just got embarrassed by the French... — original source. Check the original for full context.

Source published: 2026-10-07

Key ideas

Mistral unveiled Mistral Large 4, a 1-trillion parameter Mixture of Experts model that activates 49 billion parameters per token, outperforming Reflection AI's new $6.3 billion compute-backed Beam model on coding benchmarks within 24 hours of its debut. (0:00)
Donald Trump announced a national 'superintelligence force' aimed at preserving US frontier dominance, declaring standard artificial intelligence already obsolete in favor of artificial superintelligence. (0:00)
Real-world police interventions triggered by centralized trust-and-safety alerts on proprietary cloud models like Claude have intensified sovereign government demand for fully air-gapped, open-weight alternatives to inspect critical infrastructure. (1:36)
Chinese lab Moonshot AI attained a $50 billion valuation ahead of a planned Hong Kong IPO, with US cloud hyperscalers negotiating to host its 2.8-trillion parameter Kimiko 3 model despite intellectual property disputes with Anthropic. (0:00)

Detailed notes

## Background Published by Fireship on YouTube on October 7, 2026, this video examines recent major shakeups in the open-weight artificial intelligence landscape. According to the description, the episode breaks down aggressive competitive moves from Mistral, Reflection AI, and Moonshot AI, alongside an introduction to the agent web-browsing tool Kernel.

## TL;DR The video breaks down a chaotic week in the open-weight AI race sparked by Trump's announcement of a national superintelligence initiative and Reflection AI's release of Beam, a 501-billion-parameter model backed by billions in compute. Beam's debut was immediately challenged when Paris-based Mistral unveiled Mistral Large 4, an open-weight trillion-parameter model boasting superior benchmark claims and security auditing capabilities. At the same time, China's Moonshot AI expanded its reach with Kimiko 3 despite controversy surrounding synthetic scraping. The episode highlights why sovereign nations and privacy-conscious users increasingly avoid closed cloud models like Claude to prevent surveillance and security risks, before examining developer infrastructure from Kernel designed to let AI agents navigate anti-bot protections.

## Atmosphere & style The video delivers Fireship's signature high-velocity, satirical tech commentary, blending deadpan cynicism with rapid-fire industry analysis. It feels chaotic, witty, and playfully irreverent as it mocks the relentless corporate hype cycles, nationalist posturing, and self-reported benchmarks of the AI frontier. Beneath the dry humor and internet memes lies a pragmatic, sharp breakdown of serious geopolitical tensions and digital privacy vulnerabilities.

## Bulletin - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) Mistral unveiled Mistral Large 4, a 1-trillion parameter Mixture of Experts model that activates 49 billion parameters per token, outperforming Reflection AI's new $6.3 billion compute-backed Beam model on coding benchmarks within 24 hours of its debut. - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) Donald Trump announced a national 'superintelligence force' aimed at preserving US frontier dominance, declaring standard artificial intelligence already obsolete in favor of artificial superintelligence. - [01:36](https://www.youtube.com/watch?v=WrCjAAl9okA&t=96s) Real-world police interventions triggered by centralized trust-and-safety alerts on proprietary cloud models like Claude have intensified sovereign government demand for fully air-gapped, open-weight alternatives to inspect critical infrastructure. - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) Chinese lab Moonshot AI attained a $50 billion valuation ahead of a planned Hong Kong IPO, with US cloud hyperscalers negotiating to host its 2.8-trillion parameter Kimiko 3 model despite intellectual property disputes with Anthropic.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) Paris-based Mistral released Mistral Large 4, a trillion-parameter open-weight Mixture of Experts model featuring a 1-million-token context window and native multimodality. - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) A newly announced US 'superintelligence force' signals a geopolitical shift toward prioritizing artificial superintelligence over standard AI systems. - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) Reflection AI launched its flagship 501-billion-parameter model, Beam, after two years of development backed by $6.3 billion in compute infrastructure. - [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) Chinese AI lab Moonshot AI secured a $50 billion valuation and initiated plans for an IPO in Hong Kong. - [01:36](https://www.youtube.com/watch?v=WrCjAAl9okA&t=96s) Prompts submitted to closed proprietary APIs are vulnerable to human auditing, model retraining, and referral to law enforcement, leading to real-world criminal arrests. - [01:36](https://www.youtube.com/watch?v=WrCjAAl9okA&t=96s) Sovereign nations deploying AI to inspect domestic critical infrastructure like power grids face severe espionage risks when routing blueprints through foreign corporate APIs. - [01:36](https://www.youtube.com/watch?v=WrCjAAl9okA&t=96s) Anthropic CEO Dario Amodei publicly acknowledged the risk that foreign intelligence operatives could already be operating inside domestic frontier AI labs. - [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) Reflection AI pre-trained Beam across 24 trillion tokens in four weeks, followed by four weeks of post-training tailored for multi-step agentic workflows and tool usage. - [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) Beam remains locked behind an invite-only waitlist without released weights, leaving Google's Apache 2.0-licensed Gemma 4 as the foremost accessible US open model. - [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) Mistral Large 4 cuts inference operating costs by using sparse MoE architecture, activating only 49 billion of its 1 trillion parameters per token. - [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) Self-reported benchmarks place Mistral Large 4 18 points ahead of Beam, edging out China's GLM 5.3 in coding and reaching the global top five in software vulnerability detection. - [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) Moonshot AI's 2.8-trillion-parameter Kimiko 3 model overwhelmed GPU infrastructure upon release and prompted US cloud giants to pursue hosting deals despite scraping allegations from Anthropic. - [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) Current open-weight deployments require balancing hardware scale and legal constraints among consumer-grade Gemma 4, compute-heavy Kimiko 3, enterprise-compliant Mistral Large 4, and domestic-focused Beam. - [04:57](https://www.youtube.com/watch?v=WrCjAAl9okA&t=297s) Autonomous web agents frequently fail due to anti-bot measures, Cloudflare protection screens, and aggressive browser fingerprinting defenses on live websites. - [04:57](https://www.youtube.com/watch?v=WrCjAAl9okA&t=297s) Kernel provides open-source browser infrastructure featuring a Config Registry that automates domain-specific proxy and fingerprint configurations using continuous real-world telemetry.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) The Dumbest Week in the Open-Weight Race

While the internet was distracted being force-fed endless iterations of the viral Rumpelstiltskin video, Paris-based AI lab Mistral dropped Mistral Large 4. The release is a massive trillion-parameter open-weight model touting native multimodality and a million-token context window—hyped by Mistral as the reigning king of its class, "assuming you adopt the American ideology of just ignoring any technology coming out of China." The release arrived in the middle of what has unfolded as "the dumbest week in the history of the open-weight race."

The chaotic week kicked off when Donald Trump announced a national "superintelligence force" designed to maintain America's technological dominance at the frontier of AI, declaring on day one that standard artificial intelligence is already obsolete and the true future belongs to artificial superintelligence (SI). Following shortly behind, Reflection AI—the $25 billion startup backed by Nvidia that positions itself as "America's open frontier lab"—finally shipped its flagship model, Beam, after two years of fundraising. Boasting 501 billion parameters and backed by $6.3 billion worth of compute infrastructure, Beam entered the market benchmarking itself against an outdated Chinese model, only to be out-benchmarked 24 hours later by Mistral's surprise release. To top off the global escalation, Chinese AI lab Moonshot AI (the creators behind Kimi) closed their latest funding round at a staggering $50 billion valuation and initiated plans for a Hong Kong IPO.

- **Mistral Large 4**: Trillion-parameter open-weight Mixture of Experts (MoE) model featuring a 1-million-token context window and native multimodal capabilities. - **Trump Superintelligence Force**: A strategic frontier initiative launching with the posture that standard AI is obsolete compared to "SI" (Superintelligence). - **Reflection AI's Beam**: Billed as America's premier open alternative, launching with 501 billion parameters after two years of hype and $6.3 billion in GPU backing. - **Moonshot AI**: The Chinese developer behind Kimi, reaching a $50 billion private valuation ahead of a planned Hong Kong initial public offering.

---

[01:36](https://www.youtube.com/watch?v=WrCjAAl9okA&t=96s) Why Governments and Felons Need Open-Weight Models

The ongoing debate over proprietary versus open-weight models is no longer just an academic argument for software enthusiasts—it has become a legal and geopolitical necessity. Proprietary cloud models pose severe privacy risks for everyday end users. In San Francisco, an individual found local authorities actively tracking him down after a joking conversation with Anthropic's Claude tripped internal automated safety wires. In another recent incident in Florida, a woman was arrested only days after informing Claude that she intended to carry out an illegal act. Every prompt sent to closed proprietary data centers is scanned by automated classifiers, reviewed by internal human trust-and-safety personnel, and used downstream to train next-generation models or, in worst-case scenarios, forwarded to law enforcement.

```mermaid flowchart TD accTitle: Proprietary vs Open-Weight Data Privacy accDescr: Diagram contrasting data risks of proprietary cloud AI with self-hosted open-weight pipelines.

subgraph Proprietary["Proprietary Cloud Pipeline"] P1["User Prompt / Schematics"] --> P2["Cloud Data Center"] P2 --> P3["Safety Classifier & Human Review"] P3 --> P4["Model Retraining / Law Enforcement"] end

subgraph OpenWeight["Self-Hosted Open-Weight Pipeline"] O1["User Prompt / Schematics"] --> O2["Local Data Center / PC"] O2 --> O3["Internal Private Output"] O3 --> O4["Zero External Telemetry"] end ```

This vulnerability scales up to national security concerns. Sovereign governments looking to deploy AI to inspect critical infrastructure—such as hunting for vulnerabilities within their domestic power grids—face an impossible dilemma when using proprietary solutions like Mythos. Leveraging such models demands uploading sensitive schematics and private state blueprints to the servers of foreign corporations, leaving states vulnerable to data leaks or industrial espionage. Anthropic CEO Dario Amodei has even publicly conceded that foreign spies may already be operating inside domestic AI labs. Consequently, nations worldwide are prioritizing open-weight models that can run fully air-gapped on sovereign hardware.

- **Privacy Invasions**: Prompts sent to closed APIs undergo continuous automated filtering, classifier checks, and potential human auditing. - **Real-World Prosecutions**: Multiple legal detentions have stemmed from unredacted chat transcripts on centralized services like Claude. - **Critical Infrastructure Risks**: Inspecting power grids or defense systems via external APIs exposes classified sovereign blueprints to foreign jurisdictions and corporate insider threats.

---

[02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) The Great Model Mid-Off: Beam vs. Mistral Large 4 vs. Kimiko 3

The international race among the US, Europe, and China has produced a showdown of competing architectures and compromises. Reflection AI launched Beam backed by $4.6 billion in venture capital—raised "billion with a B" by two former DeepMind researchers arguing that the Western world needed an open-weight champion because Meta's Llama 4 was falling behind. Beam's 501-billion-parameter base model underwent an intensive four-week pretraining run on approximately 24 trillion tokens, followed by an additional four weeks of post-training optimized for complex tool use, code generation, and multi-step agentic workflows. Reflection AI claims Beam matches the performance of Zhipu AI's GLM 5.2 while utilizing only a quarter of the hardware footprint; however, Beam remains locked behind an invite-only waitlist without publicly accessible weights, meaning Google's Gemma 4 (released under an Apache 2.0 license) ironically remains America's most viable genuinely open model.

Mistral countered with Mistral Large 4, an architecture styled like "something that sounds like you'd order it from McDonald's." Using a sparse Mixture of Experts (MoE) design, the model houses roughly 1 trillion total parameters, yet activates only 49 billion parameters per token. This matches the efficiency playbook pioneered by DeepSeek to slash operational inference costs to the equivalent of running a model twenty times smaller. On public benchmarks, Mistral Large 4 narrowly edges out China's GLM 5.3 in complex coding benchmarks, outperforms Beam by 18 points, and lands in the global top five for software security vulnerability detection—the exact capability Mistral aims to monetize for state and utility infrastructure. However, the catch is that these benchmarks remain self-reported "trust me, bros" numbers; the model has not finished training, and the actual weights are scheduled for later release under a restrictive, proprietary Mistral license.

```mermaid flowchart TD accTitle: Open-Weight Model Decision Guide accDescr: Flowchart illustrating how to select an open-weight model based on hardware and regulatory constraints.

Start["Choose an Open-Weight Model"] --> Hardware{"Hardware Available?"}

Hardware -->|"Consumer Gaming PC"| Gemma["Gemma 4<br/>(Google / Apache 2.0)"] Hardware -->|"Enterprise Data Center"| Compliance{"Compliance Department?"}

Meanwhile, China's Moonshot AI already delivered on these promises months ago with Kimiko 3, a massive 2.8-trillion-parameter open model launched in July that saw demand overwhelm their GPU clusters within 48 hours. Kimiko 3 has become so influential that American cloud giants Microsoft, Amazon, and Google have actively negotiated to host it, despite Anthropic alleging Moonshot extracted Claude's capabilities by funneling over 300,000 synthetic queries through fraudulent accounts. Selecting the right model ultimately hinges on your resources and legal landscape:

- **Gemma 4**: The practical choice for local hobbyists running consumer gaming hardware ("get your money up"). - **Kimiko 3**: The premier powerhouse for teams commanding full data center infrastructure operating outside strict Western compliance mandates. - **Mistral Large 4**: The corporate enterprise choice for organizations needing high-tier security analysis alongside formal compliance guarantees. - **Beam**: The patriotic alternative for domestic institutions holding out for Reflection AI's official public weight drop.

---

[04:57](https://www.youtube.com/watch?v=WrCjAAl9okA&t=297s) Agent Web Browsing with Kernel

Regardless of which open model developers adopt, autonomous systems must interface with the live internet to perform meaningful agentic work. This workflow introduces massive friction via modern anti-bot protections, Cloudflare screens, and browser fingerprinting defenses that routinely break autonomous scripts.

Kernel addresses this challenge by providing open-source browser infrastructure built specifically for AI agents. To tackle the issue of agents getting halted by bot checks, Kernel introduced its Config Registry. Developers supply a target URL through the Kernel API or user dashboard, and the platform returns the exact proxy and browser fingerprint configurations mathematically proven to work for that specific domain. Rather than guessing headers, Kernel derives these configurations through automated, round-the-clock testing against live sites, providing users with transparent telemetry for every preset including historical success rates, sample sizes, and validation timestamps. Viewers can apply promo code `FIRESHIP` for $30 in free platform credit to unlock their first month on Kernel's Hobbyist tier.

- **Config Registry**: Automatically provisions optimal proxy configurations and browser parameters matched to specific URLs. - **Continuous Validation**: Benchmarks browser configurations against real-world bot defenses with visible success metrics and sample sizes. - **Fireship Promotion**: Using code `FIRESHIP` grants $30 in platform credits, covering a month on the Hobbyist tier.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) assuming you adopt the American ideology of just ignoring any technology coming out of China.

> [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) the dumbest week in the history of the open-weight race.

> [00:00](https://www.youtube.com/watch?v=WrCjAAl9okA&t=0s) America's open frontier lab

> [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) billion with a B

> [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) something that sounds like you'd order it from McDonald's.

> [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) trust me, bros

> [02:44](https://www.youtube.com/watch?v=WrCjAAl9okA&t=164s) get your money up

## Mentioned Amazon, Anthropic, Beam, Claude, Cloudflare, Dario Amodei, DeepMind, DeepSeek, Donald Trump, Fireship, Gemma 4, GLM 5.2, GLM 5.3, Google, Kernel, Kimi, Kimiko 3, Llama 4, Meta, Microsoft, Mistral, Mistral Large 4, Moonshot AI, Mythos, Nvidia, Reflection AI, Zhipu AI

## Video description KERNEL is the simplest way to let your agents act on any website. Use code: FIRESHIP to get a free upgrade to the Hobbyist tier ($30 value) at https://kernel.sh/fireship

Mistral, Reflection AI, and Moonshot all made big moves in the open-weight race this week. Let's dive in.

Want more Fireship?

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

[Source](https://www.youtube.com/watch?v=WrCjAAl9okA)

中文简报

Mistral力压Reflection AI的Beam领跑全球开放权重AI竞赛与专有云API相关的监控和间谍风险正促使政府和企业转向开放权重模型，尽管面临权重未发布和基准测试存疑的问题。本期视频剖析了开放权重AI竞赛中动荡混乱的一周：特朗普宣布启动国家超级智能计划，Reflection AI发布了由数十亿美元算力支持的5010亿参数模型Beam。随着总部位于巴黎的Mistral推出Mistral Large 4，Beam的亮相随即遭遇挑战；该模型是一款声称在基准测试表现和安全审计能力上均更胜一筹的万亿参数开放权重模型。与此同时，尽管围绕合成数据抓取存在争议，中国的月之暗面（Moonshot AI）仍凭借Kimiko 3扩大了影响力。本期内容重点分析了主权国家和注重隐私的用户为何日益规避像Claude这样的封闭云端模型以防范监控和安全风险，随后探讨了Kernel为协助AI代理突破反爬虫防护而打造的开发者基础设施。 Mistral推出了Mistral Large 4，这是一款拥有1万亿参数的混合专家模型，每个token激活490亿参数，在其亮相后24小时内，在编程基准测试上的表现便超越了Reflection AI由63亿美元算力支持的新模型Beam。 (0:00) 唐纳德·特朗普（Donald Trump）宣布建立一支旨在维护美国前沿主导地位的国家“超级智能部队”，并宣称传统人工智能已然过时，取而代之的是通用超级智能。 (0:00) 像Claude这样的专有云模型因集中式信任与安全警报引发了现实中的警方介入，这加剧了主权政府对完全物理隔离、用于审查关键基础设施的开放权重替代方案的需求。 (1:36) 中国实验室月之暗面（Moonshot AI）在计划赴港IPO前估值达到500亿美元，尽管与Anthropic存在知识产权争议，美国超大规模云服务商仍在谈判托管其2.8万亿参数的Kimiko 3模型。 (0:00)

Reading summary

The global open-weight AI race has escalated into a high-stakes geopolitical contest among the United States, France, and China, driven by surveillance risks and sovereign security demands. Privacy and National Security Force Sovereign Deployments (1:36) Centralized, proprietary cloud models present significant privacy and legal liabilities for end users and sovereign institutions. Prompts routed to closed data centers undergo automated classifier filtering, internal trust-and-safety audits, and potential integration into training pipelines. In real-world enforcement actions, local authorities in San Francisco pursued an individual after a joke tripped automated flags on Anthropic's Claude, and a woman in Florida was arrested days after describing an illegal act to the service. At the sovereign level, governments face severe exposure when using closed commercial models to audit critical domestic infrastructure, such as power grid vulnerabilities. Deploying external APIs requires transmitting sensitive technical blueprints to foreign corporate infrastructure, introducing espionage vulnerabilities. Anthropic CEO Dario Amodei has publicly acknowledged that foreign intelligence operatives could already be working within domestic AI laboratories. As a result, state actors and compliance-restricted organizations are increasingly prioritizing open-weight models that operate completely air-gapped on local hardware. Frontier Model Trade-Offs Across Competing Nations (2:44) The race to deliver top-tier open models has produced competing technical trade-offs between US, European, and Chinese developers. Reflection AI introduced Beam, a 501-billion-parameter base model pretrained on 24 trillion tokens and backed by $4.6 billion in venture capital and $6.3 billion in GPU infrastructure. Reflection claims Beam matches the performance of China's GLM 5.2 at one-fourth the compute footprint, yet the model remains restricted behind an invite-only waitlist without downloadable weights—leaving Google's Apache 2.0-licensed Gemma 4 as the main genuinely accessible American option. In response, France's Mistral announced Mistral Large 4, a sparse Mixture of Experts (MoE) architecture housing 1 trillion total parameters while activating only 49 billion per token to slash inference costs. Mistral claims its model outperforms Beam by 18 points, edges out China's GLM 5.3 in coding, and ranks among the global top five in software vulnerability detection. However, training is incomplete, benchmarks remain self-reported, and weights are slated for release under a restrictive proprietary license. Meanwhile, China's Moonshot AI—valued at $50 billion ahead of a planned Hong Kong IPO—previously shipped Kimiko 3, a 2.8-trillion-parameter MoE model that overwhelmed server clusters within 48 hours and drew hosting interest from major US cloud providers, despite Anthropic alleging Moonshot distilled Claude's capabilities via 300,000 synthetic queries. Infrastructure Challenges for Autonomous Agent Execution (4:57) Transitioning open-weight models into autonomous web agents introduces severe operational friction when interacting with external websites. Automated agents frequently encounter bot mitigation systems, browser fingerprinting protections, and Cloudflare challenges that disrupt execution scripts. To bridge this gap, developer tooling such as Kernel provides open-source browser infrastructure tailored for agents. Its Config Registry uses automated, round-the-clock testing against live domains to supply agents with verified proxy parameters and browser header combinations, accompanied by historical success rates and sample size telemetry. While developers market open-weight architectures as a counterbalance to proprietary monopolies, the ecosystem remains constrained by restrictive licensing, unreleased weights, and unverified vendor benchmarks.

Source quotes

[0:00] assuming you adopt the American ideology of just ignoring any technology coming out of China. [0:00] the dumbest week in the history of the open-weight race. [0:00] America's open frontier lab [2:44] billion with a B [2:44] something that sounds like you'd order it from McDonald's. [2:44] trust me, bros [2:44] get your money up

Original transcript

[0:00] Yesterday, while you were being [0:01] force-fed every iteration of that [0:02] Rumpelstiltskin video, the Paris-based [0:05] AI lab Mistral released Mistral large [0:07] four, a trillion-parameter open-weight [0:09] model that they're claiming is the best [0:11] of its kind, assuming you adopt the [0:13] American ideology of just ignoring any [0:15] technology coming out of China. It's [0:17] natively multimodal, has a million-token [0:19] context window, and claims to blah blah [0:22] blah. The more interesting part of this [0:23] release was the timing because it landed [0:26] in the middle of what many are calling [0:27] the dumbest week in the history of the [0:29] open-weight race. It kicked off early in [0:31] the week when Trump announced a [0:32] superintelligence force [0:34] >> [music] [0:37] >> whose job it is to keep America at the [0:39] frontier of the AI race, with their [0:40] first order of business claiming that AI [0:42] is actually dumb and the real future is [0:45] SI. Then Reflection AI, the $25 billion [0:48] Nvidia-backed startup that calls itself [0:50] America's open frontier lab, finally [0:52] shipped their first model after 2 years [0:54] of fundraising. It's called Beam. It has [0:57] 501 billion parameters. It was trained [0:59] on 6.3 billion dollars worth of GPUs. It [1:02] benchmarked itself against an already [1:04] outdated Chinese model. And 24 hours [1:06] later, it got out-benchmarked itself by [1:08] Mistral. Then, just yesterday, Moonshot [1:11] AI, the Chinese lab behind Kimi, closed [1:13] its final funding round at a $50 billion [1:16] valuation and started prepping a Hong [1:18] Kong IPO, which I'm told the [1:19] superintelligence force has no interest [1:21] investing in. In today's video, we'll [1:23] look at what the French, the Americans, [1:25] and the Chinese each shipped this week, [1:27] find out why every government on Earth [1:29] suddenly wants an open model to call its [1:31] own, and figure out which one you should [1:32] actually use. It is October 7th, 2026, [1:35] and you're watching The Code Report. If [1:37] you're still not convinced about the [1:38] benefits of open-weight models, last [1:40] month a man in San Francisco learned [1:42] them the hard way when he had the [1:43] authorities looking for him after a [1:45] joking conversation he had with Claude [1:47] triggered a safety review. And just last [1:49] week, a woman in Florida got arrested [1:51] just days after telling Claude she was [1:53] going to do something that's typically [1:54] frowned upon in modern society. And even [1:57] if you're not planning on committing a [1:58] felony with proprietary models, you're [2:00] still sending every prompt to a data [2:02] center, which is scanned by a [2:03] classifier, but potentially read by a [2:05] safety team, and then used in whatever [2:07] way the lab wants, like training future [2:09] models, or kicking off the process to [2:11] throw you in jail. And whole ass [2:12] countries are having the same problem, [2:14] too. If take a government that wants to [2:16] use AI to find issues in their own power [2:18] grid, using a model like Mythos, which [2:20] is designed for this exact purpose, [2:22] would require handing a foreign company [2:24] your private schematics, and then hoping [2:26] that they never get leaked, or that [2:27] someone on their payroll isn't a foreign [2:29] spy, which Dario himself has said may [2:31] already be the case. Where if you were [2:33] using an open weight model, most of [2:35] these problems go away, which is why [2:37] nearly every country in the world is [2:38] trying to build their own if they can, [2:40] or find ones they can trust if they [2:42] can't. Which brings us back to this week [2:44] and the great model mid-off. The first [2:46] announcement came from Reflection AI, [2:47] which was started by two DeepMind [2:49] researchers who raised 4.6 billion, [2:52] billion with a B, on the premise that [2:54] the best open model in the world should [2:56] come from the West, and Llama 4 wasn't [2:58] cutting it. Well, this week we finally [3:00] got to witness manifest destiny realized [3:02] when they released Beam. It's got 501 [3:04] billion parameters, and its base model [3:06] was trained in just under 4 weeks on [3:08] about 24 trillion tokens. After that, [3:11] they spent 4 more weeks post training it [3:13] to get it better at writing code, using [3:15] tools, and finishing long agent tasks. [3:17] The result is a model that Reflection [3:19] says keeps up with Jipu AI's GLM 5.2, [3:23] while needing only a quarter of the [3:24] hardware. But sadly, as of today, you [3:26] still need an invite to use it, and the [3:28] weights aren't actually out yet. Which [3:30] means the best open American model is [3:32] somehow still Gemma 4, which Google [3:34] released under Apache 2.0 back in April. [3:37] Then the next day, Mistral showed up [3:39] with something that sounds like you'd [3:40] order it from McDonald's. It's a mixture [3:42] of experts model, which means it has [3:44] about a trillion parameters, but only 49 [3:46] billion of them actually turn on for a [3:48] given token. So, the idea is that you [3:50] get a trillion parameters worth of [3:52] knowledge for roughly the cost of [3:53] running a model 20 times smaller, which [3:55] if that sounds familiar, it's probably [3:57] because it's the same trick DeepSeek [3:59] used last year to tank Nvidia. On the [4:01] TMBB's, it barely beats GLM 5.3 at long [4:04] coding jobs. It beats Beam by 18 points [4:07] and lands in the top five for finding [4:09] security bugs, which is the one Mistral [4:11] actually cares about because the power [4:12] grid scenario from earlier is the thing [4:14] they're actually trying to sell. The [4:16] catch is the numbers really are, trust [4:18] me, bros, because they were provided by [4:20] Mistral and they actually haven't [4:21] finished training yet and the weights [4:23] don't come out until later this month [4:24] under their own custom license, which [4:26] brings us to China, which already has [4:28] the thing both of them are trying to [4:29] build with Kimiko 3. It's a 2.8 trillion [4:32] parameter open model that Moonshot [4:34] released back in July, which became so [4:36] popular they had to stop taking new [4:38] subscribers 2 days later because they [4:40] ran out of GPUs. Since then, Microsoft, [4:43] Amazon, and Google have all been [4:44] negotiating to host it. Anthropic [4:46] accused them of running 300,000 requests [4:48] through fake accounts to copy Claude and [4:50] the company went from a $4 billion [4:52] valuation last November to 50 billion [4:54] yesterday with China's own state AI fund [4:56] involved in the round. So, which one [4:58] should you actually use, though? If all [4:59] you've got is a gaming PC, go with Gemma [5:01] 4 and get your money up. If you have [5:03] your own data center and no compliance [5:05] department, go with Kimiko 3. Or if you [5:07] have a data center and a compliance [5:08] department, then go with Mistral large [5:10] 4. And if you're a patriot, Beam's [5:12] weights are coming later this month. But [5:14] no matter which model you choose, [5:15] they'll need to get on the internet, [5:17] which they can do with Kernel, the [5:18] sponsor of today's video. They provide [5:20] open-source browser infrastructure that [5:22] lets your agents act on the internet [5:24] blazingly fast, but none of that [5:26] actually matters if your agent keeps [5:27] getting blocked by bot checks, which is [5:29] why they just launched config registry. [5:32] Send it the URL you want to access from [5:34] either their dashboard or the API and [5:36] it'll give you back the browser and [5:38] proxy settings that have worked for that [5:39] site. Right now, my agent keeps getting [5:41] stopped by a bot check, but when I spin [5:43] up a browser with the config Kernel [5:45] recommends, it's able to access the page [5:47] without any issues. The kernel builds [5:49] these configs by continuously testing [5:51] real browser and proxy setups against [5:53] the live site, and each one comes with a [5:55] success rate, a sample size, and last [5:57] verified date. Use the code Fireship at [5:59] the link below to get $30 in free [6:01] credits, which covers your first month [6:03] on the hobbyist plan. This has been the [6:05] Code Report. Thanks for watching, and I [6:07] will see you in the next one.