最新开放模型盘点（#23）：Laguna S2.1、Inkling 与 Kimi K3 展现开放模型在帕累托前沿的价值

整合一直是许多敏锐观察者预测模型训练实验室近期将走上的道路之一。由于训练成本每年都以数量级增长，它被贴上了「不可避免」的标签。然而，作为一个在 2024 年会预测整合将在 2026 年或 2027 年真正加速的人，我们如今身处何地？现在，更多公司正在训练强大的模型——整个投入仍然轻松达到数亿美元至数十亿美元——而且越来越多的组织正以开放方式发布这些模型。

对 token 的需求极高，并且随着模型效率提升、可解锁的使用场景增多，这种需求很可能继续增长。我们曾认为需要整合的所有这些实验室，正在意识到构建 token 机器可能是一条创造价值的路径；随着时间推移，更多公司也会发现这一价值来源。

Thinking Machines 是最典型的例子：2025 年 2 月该公司宣布成立时，包括我在内，很少有人会将其归入开放模型公司。如今，它的开放模型微调服务每年已带来数亿美元收入，而且正在发布美国构建的最佳开放权重模型——领先于 NVIDIA 的 Nemotron 和 Arcee 的 Trilogy 等早期领跑者。

分享

在生态系统的另一端，中国实验室仍保持着持续发布的节奏；小米等新进入者也仍在更广泛的 AI 经济中积累影响力。长期以来我一直预测行业会整合，但如今更稳妥的判断似乎是持续采用，并尝试想象开放模型在其中扮演的角色。Kimi K3 这类营收分成许可证能持续多久？开放模型能够获得多大市场份额？我们正在进入决定性的时代。

这是我们有史以来内容最密集的一期开放模型盘点之一，我们很兴奋！

• Inkling，来自 thinkingmachines：Thinking Machines 的首个模型，是一个 975B-A41B 多模态 MoE，支持文本、图像和音频输入，并输出文本。虽然在同体量级别的同类模型中（尤其是中国模型中），它并不是最强者，但它被定位为极佳的微调基础模型，例如可通过其商业产品 Tinker 进行微调。该公司还发布了更小的版本（276B-A12B），其在同等规模中极具竞争力。

• Hy3，来自 tencent：腾讯推出的一个 295B-A21B MoE。它在所有指标上均优于前代产品。不过最值得注意的是许可证的变化：前一版本（在第 21 期 Artifacts中介绍）采用自定义且限制颇多的许可证，而腾讯此次发布改用了 Apache 2。该模型还能够证明一道有 50 年历史的数学问题（使用专门的测试框架，并由 Sol 担任评判；尽管后者究竟有多重要并不清楚）。

• Laguna-S-2.1，来自 poolside：Poolside 很快从默默无闻跃升为 Artifacts 的常客，这是它连续第三个月登场。S2.1 是 118B-A8B MoE 的新版预训练和后训练版本，可装入一台 DGX Spark，这为其带来了大量关注。Poolside 还采用了 OpenMDW 许可证：这是一种类似 Apache 2.0 的免费许可证，但针对 AI 模型具备更完善的法律支持。该公司还在博客中提供了更多细节，其中包括全部评估轨迹。对于一次开放模型发布而言，这提供了大量透明度！

• DeepSeek-V4-Flash-0731，来自 deepseek-ai：就在 OpenAI 将其最小模型价格下调 80%的一天后，这头巨鲸发布了 V4 Flash 模型更新，在帕累托前沿上击败了 Luna。更大的模型尚未更新，因此其最终性能落点仍有待观察。在最初的 V4 发布中，Flash 版本在单位参数性能方面是主角，而 Pro 的表现则相当令人失望。

• Kimi-K3，来自 moonshotai：这是近期规模最大的开放模型发布之一，我们已在单独的文章和一期播客节目中进行介绍。它采用非商业许可证发布，要求推理和微调服务提供商签订商业协议。Kevin Xu 与 Graham Webster 在一篇文章中认为，这类许可证可能使美国政府未来能够对与中国 AI 公司开展业务的美国实体采取行动：

但如果一家美国公司需要与 Moonshot 签订合同，才能提供 Kimi K3 生成所需的推理 token，情况就不同了。美国官员和其他人士曾讨论过的一些政策工具，作为限制使用中国开放模型的潜在杠杆，在这种情况下将更明确地适用。

• LongCat-2.0，来自 meituan-longcat：这家中国版 DoorDash 再次回归。这次，公司发布了另一款拥有 1.6T 参数的大型 MoE。尽管该模型本身除基准测试外，在同等规模中并非能力最强，但它完全使用 Ascend 910 训练，使其成为首个完全在中国加速器上训练的非华为、非玩具级模型。其他中国芯片大多用于推理（如果有使用的话）。

• Laguna-XS-2.1，来自 poolside：Poolside 小型（33B-A3B）MoE 的更新版本。

• Motif-3-Beta，来自 Motif-Technologies：韩国公司 Motif 推出的 314B-A13B MoE 预览版。这是该公司迄今最具雄心的模型：其规模显著更大，并引入了 GDLA 和 mHC 等架构创新。

• Apertus-v1.5-70B，来自 swiss-ai：在完全开源的 Apertus 1.0 基础上继续预训练，额外使用了 2T token。

• Instella-MoE-16B-A3B-Think，来自 amd：AMD 使用 Instinct 加速卡训练的一款 16B-A3B MoE。AMD 还提供了从基础模型到 SFT 检查点的所有不同阶段，以及 MidTrain 和 DPO。

Consolidation has been one of the paths that many astute observers predicted for the near-future of labs training models. It was labelled as inevitable, as training costs are increasing by orders of magnitude every year. Yet, as someone who in 2024 would’ve predicted consolidation really picking up come 2026 or 2027, where are we? We’re at a place where more companies are training strong models — easily investing hundreds of millions to billions of dollars in the total effort still — and an increasing number of organizations are releasing these models openly.

The demand for tokens is incredibly high, and likely to increase as models get more efficient and unlock more possible use cases. All of these labs we thought would need to consolidate are realizing that building token machines is a likely path to value, and more companies will identify that source of value over time.

The prime example is Thinking Machines — when they announced their company in February 2025, very few people would’ve put them in the bucket of an open models company, myself included. Now their open model finetuning service is making hundreds of millions in revenue per year and they’re releasing the best open-weight models built in the U.S.A. — ahead of the early leaders in NVIDIA with Nemotron and Arcee’s Trilogy.

Share

On the other side of the ecosystem is the sustained pace from the Chinese labs, with newer entrants like Xiaomi still accumulating mindshare in the broader AI economy. Having predicted consolidation for a long time, it now seems like a safer bet is to predict continued adoption, and try to imagine the role that open models play there. How much can revenue-share licenses like Kimi K3 stick? How much market share can open models take? We’re entering the decisive era.

This is one of the most packed recaps of open models we’ve ever had, we’re excited!

• Inkling by thinkingmachines: The first model from Thinking Machines is a 975B-A41B multimodal MoE that supports text, images, and audio as inputs and produces text as output. While it is not the strongest model among peers (in China) in its size class, it is positioned to be a great base for fine-tuning, e.g., through their commercial offering, Tinker. They also release a smaller version (276B-A12B), which is really competitive for its size.

• Hy3 by tencent: A 295B-A21B MoE from Tencent. It improves over its predecessor across all metrics. Most notable, however, is the license change: While the previous version (covered in Artifacts 21) used a custom and rather restrictive license, Tencent switched to Apache 2 for this release. The model was also able to proof a 50 year old math problem (with a dedicated harness and Sol as a judge, although it is unclear how important the latter really is).

• Laguna-S-2.1 by poolside: Poolside quickly rose out of nowhere to become a frequent guest at Artifacts, marking its third appearance in three consecutive months. S2.1 is a newly pre- and post-trained version of the 118B-A8B MoE that fits on a DGX Spark, which brought it a lot of attention. Poolside also adopted the OpenMDW license, which is an Apache 2.0-like free license but has better legal backing for AI models specifically. The company also goes into more detail in its blog, which includes all the evaluation trajectories. This is a lot of transparency for an open model release!

• DeepSeek-V4-Flash-0731 by deepseek-ai: Just one day after OpenAI has dropped the prices of their smallest model by 80%, the whale dropped an update to their V4 Flash model, beating Luna at the pareto frontier. The bigger model is not updated yet, so it remains to be seen where it will land in terms of performance. For the initial V4 releases, the Flash version was the star of the show in terms of performance per parameter, while Pro was rather underwhelming.

• Kimi-K3 by moonshotai: This is the biggest open model release in some time, and we covered it in a separate post and a podcast episode. It was released under a noncommercial license, requiring inference and fine-tuning providers to enter into a commercial agreement. Kevin Xu and Graham Webster argue in a post that these licenses enable potential future government action against US entities doing business with Chinese AI companies:

But if a US company needs a contract with Moonshot to provide the inference tokens that Kimi K3 generates, the picture looks different. Some of the policy tools US officials and others have debated as potential levers to restrict Chinese open model use would more clearly apply.

• LongCat-2.0 by meituan-longcat: The Chinese DoorDash is back again. This time, the company released another big MoE with 1.6T parameters. While the model itself is not the most capable for its size beyond benchmarks, it was trained entirely on Ascend 910s, making it the first non-Huawei, non-toy model trained entirely on Chinese accelerators. Other Chinese chips are mostly used for inference (if at all).

• Laguna-XS-2.1 by poolside: An update to the small (33B-A3B) MoE from Poolside.

• Motif-3-Beta by Motif-Technologies: A preview of a 314B-A13B MoE by the Korean Motif. This is by far the company’s most ambitious model, as it is considerably larger and introduces some architectural innovations like GDLA and mHC.

• Apertus-v1.5-70B by swiss-ai: A continued pre-train of the fully open-source Apertus 1.0, using 2T more tokens.

• Instella-MoE-16B-A3B-Think by amd: A 16B-A3B MoE trained by AMD on Instinct cards. AMD also provides all the different stages, from the base to the SFT checkpoints, as well as MidTrain and DPO.