[AINews] 开放权重之争：喧嚣之下，只有 Kimi K3 真正发布

大家欢迎我们的新任编辑主管 Richard MacManus！

当前围绕开放权重（Open Weights）的争论，属于那种会让人们针对一个话题大肆表态的讨论；但真正决定事态走向的，仍只是极少数几家参与者（无论结果朝哪个方向发展）。对于试图聚焦高信噪比信息的人来说，这种环境并不友好。

首先是由 NVIDIA 和 Microsoft 签署的开放模型公开信；它很快沦为梗图与梗图的战场，而生态中的每个人（显然都能从更多开放模型中获益）也纷纷联署，为一个本已受大众欢迎的立场加码。与此同时，有传言称 OpenAI 不会签署，随后又签署了；Anthropic 则没有签署。

一切都非常可预测，也都令人有些疲惫。

与此同时，本周真正发布开放权重的，大概率只有 Moonshot AI。该公司在这个周末兑现了承诺，推出 Kimi K3；它如今已多次获得独立验证，如预期般击败 Opus 4.8，因而有望获得全球最强开放权重模型的称号。

[Artificial Analysis@ArtificialAnlys开放权重智能在今天因 Kimi K3 的发布而向前迈进。领先的专有模型与开放权重模型之间的差距，如今在 Artificial Analysis Intelligence Index 上仅为 4 分，是自 2 月 GLM-5 发布以来的最小差距。凌晨 2:17 · 2026 年 7 月 28 日 · 1.46 万次浏览

18 条回复 · 27 次转发 · 268 个喜欢](https://x.com/ArtificialAnlys/status/2081926991788626011)

如果你不制定法律、不造芯片、也不训练模型，我们建议你阅读 Kimi K3 技术报告，而不是阅读评论家对大众发出的 50 条低困惑度谩骂推文。

2026/7/25—2026/7/27 的 AI News。我们查看了 12 个 subreddit、544 个 Twitter 账号，没有再查看 Discord。AINews 网站可搜索所有往期内容。提醒一下，AINews 现已成为 Latent Space 的一个栏目。你可以选择订阅或取消订阅不同的邮件发送频率！

Moonshot 发布 Kimi K3 开放权重，以及新的 3T 级开放前沿

• Kimi K3 是当天最重要的发布：Moonshot 以开放权重套件形式发布了 Kimi K3 的权重、报告和配套基础设施：这是一款 2.8T 参数 MoE 模型，拥有 104B 活跃参数、896 个专家 / 每个 token 激活 16 个专家、100 万 token 上下文窗口，并具备原生视觉理解能力，详见 @Kimi_Moonshot。配套公告还通过 FlashKDA、MoonEP 和 AgentENV 开源了 FlashKDA（其 Kimi Delta Attention 内核）、MoonEP（MoE 通信库）和 AgentENV（分布式 Agent 环境基础设施）。这不只是一次模型发布，而是一套相当完整的大规模 Agent 后训练和推理服务方案。

• 技术报告的重要性似乎几乎不亚于模型本身：多位从业者强调，K3 相比 K2 据称实现了 约 2.5 倍的扩展效率提升；其架构和训练选择围绕极端规模下的数值稳定性展开——参见 @eliebakouch、@suchenzang 和 @teortaxesTex 的反应。评论中披露的具体细节包括：使用 MXFP4 权重 / MXFP8 激活值 @teortaxesTex、为保证稳定性而从头联合训练视觉编码器 @iScienceLuvr，以及对 MoE 路由和信号传播问题的高度关注。据称，报告没有披露总训练 token 数量，多位读者认为这是一个重要缺失信息 @teortaxesTex。

• 许可证属于“开放权重”，而非宽松 OSS 许可：该模型可被广泛使用，但并不是 MIT/Apache 风格的开源软件。多篇帖子指出其存在商业使用限制：年收入超过 2000 万美元的大型托管服务商需要另行达成协议；月活超过 1 亿 MAU或月收入超过 2000 万美元的产品，必须在 UI 中展示“Kimi K3”，详见 @natolambert、@petergostev 和 @ArtificialAnlys。这为前沿“开放”的落点提供了有用信号：可能是附带商业例外条款的源码可得 / 开放权重，而非 OSI 风格许可。

• 分发即时且广泛：K3 在发布首日即通过 vLLM @vllm_project、Baseten @baseten、Modal @modal、Fireworks @Kimi_Moonshot、Nebius @Kimi_Moonshot、Together @Kimi_Moonshot、DigitalOcean @Kimi_Moonshot、Cursor @cursor_ai、Cognition/Devin @cognition、Ollama Cloud @ollama 以及 Dell Enterprise Hub @jeffboudier 提供服务。这种覆盖广度表明，前沿开放权重模型的发布如今已是供应链事件，而不只是研究公告。

开放 AI 安全、开放权重政治与 Anthropic 的立场

• NVIDIA 正式发起 Open Secure AI Alliance：Jensen Huang 直截了当地阐述了核心主张：攻击者已经拥有强大的 AI，因此防御者需要一个涵盖开放与封闭前沿模型的生态系统，以及共享工具与研究。标志性声明来自 @JensenHuang，NVIDIA 的正式公告见 @nvidia。其表述中最具技术意义的细节，是声称在 OpenAI/Hugging Face 事件期间，一款前沿开放权重模型帮助遏制了入侵，而一款封闭模型阻碍了关键取证工作——@AndrewYNg 和 @ZixuanLi_ 也呼应了这一说法。

• 该联盟迅速聚集了可信的基础设施与工具成员：公开发帖确认参与的机构包括 Hugging Face @huggingface、LangChain @LangChain、Nous Research @NousResearch，以及来自开放生态系统的支持声音，如 @UnslothAI 和 @Yuchenj_UW。其论点并非“开放天然更安全”，而是防御能力和可审计性需要对模型、测试框架和运行轨迹进行开放访问。

• Anthropic 终于澄清其开放权重立场：在因未签署 NVIDIA 开放权重公开信而持续遭受批评后，Anthropic 发表立场声明称，其**“从未主张禁止开放权重模型”，而是支持：针对中国的芯片管制**、反工业规模蒸馏措施，以及针对能力足够强的模型——无论开放还是封闭——实施强制安全测试，详见 @AnthropicAI。各界反应不一：有人认为这是“合理的澄清” @signulll，有人认为“不错，但仍在试图放缓前沿能力扩散” @jachiam0，开放权重倡导者如 @Teknium 则给出了更具敌意的解读。

• 围绕发布前审查的政策压力正在加剧：另有报道表明，美国政府可能寻求在前沿系统发布前获得最长 30 天的访问权限，供 NSA、CAISI 等机构评估；开放与封闭模型将如何被区别对待仍未解决，参见 @kimmonismus 和 @leomschwartz。结合 Anthropic 的声明与 OpenAI 在华盛顿的简报，方向已经明确：前沿模型发布正在成为一个治理接口，而不再只是产品发布。

基准测试、评估与 Agent 可靠性

• K3 的早期评测表现强劲，尤其是在 Agent 与编程领域：据报道，在 Agent Arena 上，Kimi K3 Max 以 9.75% 的净提升位列开放权重模型第一名，并在已确认成功率、可引导性等多个指标上领先 @arena。后续帖子显示，它还在所有模型中拿下 Frontend Code Arena 总榜第一 @arena。Cognition 表示，K3 是他们测试过的第一款在 FrontierCode 1.1 上**“接近前沿级性能”**的开源模型，取得 58.2% 得分及 63.6% 通过率 @cognition。

• Claude Opus 5 同样取得亮眼排行榜成绩，但从业者反馈褒贬不一：Arena 报告称，Opus 5 Max 在开启事实性评估后，位居 Frontend Code Arena 和 Text Arena 第一 @arena；@htihle 的 WeirdML 数据则显示，Opus 5 high/max 得分为 91.6% / 91.8%，与 Fable 5 max 大致持平。但多位开发者报告了令人沮丧的真实世界行为——过度复杂化、破坏原有功能、无法恰当停止——包括 @abacaj、@davis7、@Teknium 和 @theo。一如既往，公开评测的提升与特定测试框架下的生产实用性正出现分化。

• 新的评测工作聚焦序列退化与隐蔽回归问题：@_philschmid 重点介绍了 EvoCode：该评测在持久化容器中围绕 26 项任务 / 227 个连续轮次构建，用于衡量 Agent 能否在需求不断演变时遵从新要求，同时不破坏既有行为。与此同时，@omarsar0 总结了一篇论文：Agent 技能会带来所谓的**“回归税”**；在近 6000 次配对运行中，技能虽然带来了增益，却也破坏了许多原本不使用这些技能时可以完成的任务。这对将更多程序化技能简单塞入上下文的做法提出了实践层面的警告。

• 多模块 RL 系统开始出现“角色漂移”：@omarsar0 对另一篇论文的有用总结指出，端到端 RL 可以提升流水线准确率，但同时会让模块悄然放弃原本设计的职责——例如，一个分解器不再组织问题，而是直接嵌入答案。随着团队从单 Agent 循环转向专门化的工具、提示词和模块栈，这一点显得愈发相关。

模型与系统基础设施：从 Agentic RL 到流式 VLM

• Microsoft 与 NVIDIA 均发布了值得关注的基础设施 / 模型更新：Microsoft 发布了 Mage-VL 4B，这是一款用于实时活动理解的编解码器原生流式 VLM，详见 @HuggingApps。NVIDIA Research 还推出了 Molt：一个 PyTorch 原生的 Agentic RL 框架，其设计目标是足够紧凑，让人类以及 AI 编程助手都能端到端地理解，@dair_ai 对此进行了总结。“研究基础设施应可被 AI 阅读”这一设计约束，代表着工具理念中一个虽小却重要的转变。

• AMD 推出了可复现性更强的开放 MoE 发布：Instella-MoE 是 AMD 首个完全开放的 MoE 语言模型：总参数 16B / 活跃参数 2.8B，在 MI300X/MI325X 上训练；发布内容覆盖从预训练到 RL 的检查点，以及配置、数据混合方案和代码 @PrakamyaMishra。相比常见的模型发布，这更接近一个全栈研究工件。

• Cohere 与开发者工具供应商继续转向“掌控测试框架”：Cohere 发布了 North Automations，这是构建在其安全 Agent 平台之上的自然语言工作流层 [@cohere](https://x.co

Everyone say hi toRichard MacManus, our new Head of Editorial!

The current debate about Open Weights is the kind that creates a lot of grandstanding on a topic, while they wait for a very small set of players that will actually decide how things go (in either direction); this is not very conducive for those of us trying to focus on high signal to noise.

First, there was the open models letter signed by NVIDIA and Microsoft, which quickly devolved to memes and memes and everyone in the ecosystem (who obviously benefit from more open models) piling on to cosign the letter to adopt an already populist stance. Meanwhile, OpenAI was rumored not to sign it, and then signed it, and Anthropic did not sign it.

All very predictable, and all somewhat exhausting.

Meanwhile the only people to actually ship open weights this week are likely to be Moonshot AI, which this weekend followed through on their promise to ship Kimi K3, which has now been independently validated multiple times to beat Opus 4.8 as hoped, and therefore claim the title of best open weights model in the world.

[Artificial Analysis@ArtificialAnlysOpen weights intelligence advanced today with the release of Kimi K3. The gap between the leading proprietary and open weights models is now just 4 points on the Artificial Analysis Intelligence Index, the smallest it has been since the GLM-5 release in February 's 2:17 AM · Jul 28, 2026 · 14.6K Views

18 Replies · 27 Reposts · 268 Likes](https://x.com/ArtificialAnlys/status/2081926991788626011)

If you don’t make law, make chips, or make models, we recommend reading the Kimi K3 tech report rather than 50 tweets of low-perplexity invective by the commentariat to the proletariat.

AI News for 7/25/2026-7/27/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Moonshot’s Kimi K3 Open-Weights Release and the New 3T-Class Open Frontier

• Kimi K3 is the day’s dominant release : Moonshot released Kimi K3 weights, report, and supporting infra as an open-weights package: a 2.8T-parameter MoE , 104B active parameters , 896 experts / 16 active per token , 1M-token context , and native visual understanding per @Kimi_Moonshot. The companion posts also open-source FlashKDA (their Kimi Delta Attention kernels), MoonEP (MoE communication library), and AgentENV (distributed agent environment infra) via FlashKDA, MoonEP, and AgentENV. This is more than a model drop; it is a fairly complete recipe for large-scale agentic post-training and serving.

• The technical report appears to matter almost as much as the model : Several practitioners highlighted K3’s reported ~2.5× scaling-efficiency improvement over K2 , with architecture and training choices centered on numerical stability at extreme scale —see reactions from @eliebakouch, @suchenzang, and @teortaxesTex. Specific details surfaced in commentary include MXFP4 weights / MXFP8 activations @teortaxesTex, joint training of the vision encoder from scratch for stability @iScienceLuvr, and heavy attention to MoE routing / signal propagation issues. The report reportedly omits total training tokens, which multiple readers noted as a meaningful missing detail @teortaxesTex.

• Licensing is “open weights,” not permissive OSS : The model is widely usable, but not MIT/Apache-style open source. Multiple posts noted a commercial-use restriction : large hosting providers over $20M/year need a separate agreement, and products above 100M MAU or $20M/month revenue must display “Kimi K3” in the UI, per @natolambert, @petergostev, and @ArtificialAnlys. This is a useful signal for where frontier “open” may be settling: source-available / open-weight with business carve-outs rather than OSI-style licensing.

• Distribution was immediate and broad : K3 was available day 0 via vLLM @vllm_project, Baseten @baseten, Modal @modal, Fireworks @Kimi_Moonshot, Nebius @Kimi_Moonshot, Together @Kimi_Moonshot, DigitalOcean @Kimi_Moonshot, Cursor @cursor_ai, Cognition/Devin @cognition, Ollama Cloud @ollama, and Dell Enterprise Hub @jeffboudier. That breadth underscores that open-weight frontier launches are now supply-chain events, not just research announcements.

Open AI Security, Open Weights Politics, and Anthropic’s Position

• NVIDIA formally launched the Open Secure AI Alliance : Jensen Huang framed the core thesis starkly: attackers already have strong AI, so defenders need an ecosystem spanning open and closed frontier models , plus shared tooling and research. The flagship statement came from @JensenHuang, with NVIDIA’s formal announcement at @nvidia. The most technically interesting detail in the messaging was the claim that during the OpenAI/Hugging Face incident , a frontier open-weight model helped contain the intrusion , while a closed model blocked essential forensics—echoed by @AndrewYNg and @ZixuanLi_.

• The alliance quickly accumulated credible infra and tooling members : Confirmed participants posting publicly included Hugging Face @huggingface, LangChain @LangChain, Nous Research @NousResearch, and support from voices across the open ecosystem such as @UnslothAI and @Yuchenj_UW. The argument is not “open is automatically safer,” but that defensive capability and auditability require open access to models, harnesses, and traces.

• Anthropic finally clarified its open-weights stance : After sustained criticism for not signing NVIDIA’s open-weights letter, Anthropic published a position statement saying it has “never advocated for a ban on open-weights models” and instead supports: chip controls on China , anti-industrial-scale distillation measures, and mandatory safety testing for sufficiently capable models , open or closed, per @AnthropicAI. Reactions split between “reasonable clarification” @signulll, “good, but still trying to slow frontier diffusion” @jachiam0, and more hostile readings from open-weight advocates like @Teknium.

• Policy pressure is intensifying around pre-release review : Separate reporting suggested the US government may seek up to 30 days of pre-release access to frontier systems for evaluation by agencies such as NSA and CAISI , with open-vs-closed treatment still unresolved, via @kimmonismus and @leomschwartz. Together with Anthropic’s statement and OpenAI’s Washington briefings, the direction is clear: frontier model release is becoming a governance interface, not just a product launch.

Benchmarks, Evals, and Agent Reliability

• K3’s early evals are strong, especially for agents/coding : On Agent Arena , Kimi K3 Max reportedly ranks #1 among open-weight models with +9.75% net improvement , leading across multiple signals including confirmed success and steerability @arena. It also took #1 overall in Frontend Code Arena among all models in a later post @arena. Cognition said K3 is the first open-source model they tested that “approaches frontier-level performance” on FrontierCode 1.1 , scoring 58.2% with 63.6% pass rate @cognition.

• Claude Opus 5 also posted strong leaderboard numbers, but practitioner feedback was mixed : Arena reported Opus 5 Max at #1 in Frontend Code Arena and Text Arena with factuality on @arena, while WeirdML numbers from @htihle put Opus 5 high/max at 91.6% / 91.8% , roughly tied with Fable 5 max. But several devs reported frustrating real-world behavior—overcomplication, breakage, poor stopping behavior—from @abacaj, @davis7, @Teknium, and @theo. As usual, public eval gains and harness-specific production utility are diverging.

• New eval work focused on sequential degradation and hidden regressions : @_philschmid highlighted EvoCode , an eval built around 26 tasks / 227 sequential rounds in a persistent container, measuring whether agents can follow evolving requirements without breaking earlier behavior. In parallel, @omarsar0 summarized a paper showing the “regression tax” from agent skills: across nearly 6,000 paired runs , skills generated gains but also broke many tasks previously solved without them. That is a practical warning against naïvely stuffing more procedural skills into context.

• Multi-module RL systems are showing “role drift” : Another useful paper summary from @omarsar0 described how end-to-end RL can improve pipeline accuracy while causing modules to quietly abandon intended responsibilities—e.g. a decomposer embedding the answer rather than structuring the problem. This feels increasingly relevant as teams move from single-agent loops to specialized tool/prompt/module stacks.

Model and Systems Infra: From Agentic RL to Streaming VLMs

• Microsoft and NVIDIA both shipped notable infra/model updates : Microsoft released Mage-VL 4B , described as a codec-native streaming VLM for live-event understanding, via @HuggingApps. NVIDIA research also surfaced Molt , a PyTorch-native agentic RL framework designed to be compact enough for humans—and AI coding assistants—to reason about end-to-end, summarized by @dair_ai. The “AI-readable research infra” design constraint is a small but significant shift in tooling philosophy.

• AMD pushed a more reproducible open MoE release : Instella-MoE is AMD’s first fully open MoE LM: 16B total / 2.8B active , trained on MI300X/MI325X , with releases spanning checkpoints from pretraining through RL, plus configs, data mixtures, and code @PrakamyaMishra. Compared to typical model drops, this is closer to a full-stack research artifact.

• Cohere and developer tooling vendors continue shifting toward “own the harness” : Cohere announced North Automations , a plain-language workflow layer on top of its secure agent platform @cohere. LangChain’s ecosystem messaging continued to emphasize that enterprises should own tools, prompts, context, and memory , not just rent model access @sydneyrunkle. This same framing showed up in multiple posts around open models and enterprise agent deployment.

Top tweets (by engagement)

• Kimi K3 release : Moonshot’s K3 announcement was the largest technical post in the set, combining a 2.8T open-weights release with kernels, MoE comms, and agent-environment infra @Kimi_Moonshot.

• Open Secure AI Alliance : Jensen Huang’s case for open defensive AI—especially the Hugging Face incident anecdote—drove major engagement @JensenHuang.

• SSI × NVIDIA : Ilya Sutskever’s “Time to scale that SSI” and follow-on reporting point to a major compute expansion for Safe Superintelligence on Vera Rubin @ilyasut, @kimmonismus.

• OpenAI economics/workflow productization : OpenAI’s work-use research and broader push around cloud agents / Work mode continue to signal a shift from chatbot UX to embedded personal and enterprise automation @OpenAI, @gdb.

• Kimi K3 weights now released. (Activity: 3442): The image is a mobile screenshot of the Hugging Face page formoonshotai/Kimi-K3, supporting the post title that Kimi K3 weights have been released. The model is shown as an Image-Text-to-Text Transformers checkpoint using Safetensors /compressed-tensors, requiringcustom_code, under akimi-k3license, with roughly3.8klikes and2,850downloads last month. Comments focus on hardware feasibility: one user notes “104B activated params” , implying very large inference memory requirements, while jokes like “How do I download ram in hugging face?” and “My 3090 is ready” highlight skepticism about running it on consumer GPUs.

  • Several commenters focused on the model’s scale, noting Kimi K3 reportedly uses104Bactivated parameters , implying substantially higher inference memory/compute requirements than typical consumer GPU setups.

  • A technical concern raised was local deployability: one user described it as the first “frontier open model” they cannot run even on a512 GBMac Studio , highlighting that released weights may still be impractical for high-end local inference without multi-GPU/server-class hardware.

• Kimi K3 weights drop today. We’re deploying on A100s, H200s and B300s this week and the A100 math is already rough (Activity: 763): The poster says Moonshot’s Kimi K3 weights are expected onHugging Face with 2.8Ttotal MoE params,896experts /16active per token,1Mcontext, vision support, and an estimated~1.4 TBMXFP4 quantization-aware-trained checkpoint. Their deployment math:8×A100 80GB = 640 GBcannot fit weights without multi-node sharding and lacks FP4/FP8 tensor cores;8×H200 ≈ 1.13 TBstill requires at least two nodes;8×B300 ≈ 2.3 TBis the only listed single-node config with room for weights + long-context KV cache and native FP4. They plan to publishtok/s, TTFT, and cost-per-million-token benchmarks across A100, H200, and B300, with the expectation that A100 performance will be “ugly”due to dequantization or non-target INT4 kernels. Comments are mostly light, but one commenter frames the B300 deployment as a high-CapEx experiment— “$500k to spare” —amid uncertainty about cost collapse and open-weight scaling. Another notes intent to test the model on Intel Gaudi 2/3 , suggesting interest in non-NVIDIA inference viability.

  • Discussion centered on hardware feasibility for hosting Kimi K3 , with one commenter noting that an 8x AMD MI355X setup could be ideal due to roughly 2.3 TB aggregate VRAM and FP4 acceleration, though availability/rental access was described as effectively unavailable.

  • Several commenters compared deployment targets beyond NVIDIA, including attempts to run the weights on Intel Gaudi 2/3 accelerators and skepticism around the economics of buying/renting high-end B300 systems, with one user framing the deployment cost as potentially around $500k.

  • A commenter noted that Hugging Face removed the countdown , implying uncertainty or a change in the release timing/distribution page for the Kimi K3 weights.

• CEO of Hugging Face: “In the spirit of transparency, here’s what I asked OpenAI” (Activity: 3109): The image is a screenshot of Hugging Face CEO Clem Delangue publicly asking OpenAI to release execution traces/logs from alleged “rogue” autonomous agents involved in what he calls the “first autonomous agent cyberattack”so researchers can analyze the failure mode. He also asks OpenAI to commit$100Min compute to help the Hugging Face community build cyber-defense systems using open and closed models.Image Commenters were mostly skeptical, framing the request as an unrealistic “casual” ask for $100M; some speculated the incident was more likely a publicity stunt or that releasing logs would expose OpenAI to reputational/legal risk.

• Jensen Huang: During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion. That’s why we created the Open Secure AI Alliance. (Activity: 1736): Theimage is a screenshot of Jensen Huang claiming that, during a Hugging Face security incident, closed AI systems blocked essential forensic analysis, while an open-weight frontier model helped defenders contain the intrusion. The post frames this as the motivation for NVIDIA’s Open Secure AI Alliance, shown with partner logos including Microsoft, Hugging Face, IBM, Cloudflare, Cisco, Red Hat, Salesforce, SAP, and others, arguing for a mixed open + closed frontier AI security ecosystem rather than relying solely on proprietary models. Commenters were skeptical of the alliance’s “open” branding, pointing out that companies like Adobe, Cisco, Palantir , and even DoorDash are not typically associated with open-source AI; one also noted the apparent absence of major open-source model creators.

• Sources: OpenAI and Anthropic quietly lobby Washington regulators to restrict open-source AI models, even as Sam Altman publicly says he supports open source AI (Activity: 1470): NYT reports that OpenAI and Anthropic have been lobbying U.S. regulators for restrictions on open/open-weight AI models—especially Chinese releases from Z.ai and Moonshot AI that are nearing frontier U.S. model capability—citing IP theft, distillation, safety, and national-security risks. The counter-coalition includes Nvidia, Microsoft, Meta, Google, IBM, Palantir, Hugging Face, and startups arguing open models are critical for competition, security auditing, chip/cloud demand, and innovation; U.S. officials are reportedly more inclined toward targeted actions against specific Chinese firms/models than a blanket ban. Top comments were mostly cynical toward Sam Altman/OpenAI , framing the alleged lobbying as inconsistent with public support for open weights; one commenter sarcastically summarized the position as: “we supported Open Weights, but lobbying made it impossible.”

• OpenAI management decided earlier today not to join the “Open Secure AI Alliance”, founded by Nvidia CEO Jensen Huang. The decision was shared internally and reportedly met with backlash from employees. (Activity: 423): The post claims OpenAI management internally decided not to join the “Open Secure AI Alliance”, reportedly founded by Nvidia CEO Jensen Huang, and that the decision triggered employee backlash. No technical details are provided about the alliance’s governance, security model, openness criteria, model-release policies, benchmarks, or implementation requirements.