Back to home

Compare

Comparing: K2 Horizon Fully Open-Sources 375B Models — Big Tech's Pseudo-Open Challenged & K2 Horizon 把 375B 模型全套开源 — 大厂「伪开源」遇到对手了

AEN
K2 HorizonIFMDeepSeek·

K2 Horizon Fully Open-Sources 375B Models — Big Tech's Pseudo-Open Challenged

Last week, a team of fewer than 30 people at IFM open-sourced six models (0.9B to 375B) in a single drop — the most thorough open-source LLM release of 2025, and the most direct challenge yet to the "weights-only" playbook favored by DeepSeek and Llama.

What this is

IFM (Institute of Foundation Models) is a small California-based research outfit focused on "open and independent" frontier models. K2 Horizon is the model matrix it released in one go: six models spanning mobile (0.9B) up to a 375B flagship. Crucially, this isn't a Hugging Face dump of model files — IFM has published the training data recipe, training code, intermediate weight checkpoints (snapshots saved mid-training), every training log, and the full evaluation results.

Compare that with the rest: DeepSeek releases weights but keeps its training data private; Meta's Llama comes with commercial-use restrictions. K2 Horizon comes closer to academia's "fully reproducible" standard.

Industry view

Supporters call it a direct counter to Big Tech's "pseudo-open-source." As models grow more capable, closed-source vendors tighten their grip on developers — a genuinely open model at least gives academia and smaller companies an exit ramp.

Critics are equally sharp. First, the inference and fine-tuning costs of a 375B model are out of reach for typical teams; "fully open" does little for real-world usability. Second, the Stable Diffusion story has already played out once: hype fades, follow-up maintenance and iteration lag, and closed-source services catch up. Third, IFM has fewer than 30 people, with undisclosed funding and compute sources — whether it survives the next training run is a real question.

Impact on regular people

For enterprise IT: No short-term procurement value, but worth tracking as a technology reserve — especially the 0.9B tier, which may eventually run on edge devices inside your company.

For working professionals: No need to download anything today. Note the name for now; if over the next year you hear more and more companies talking about "private deployment," the technical choices behind it likely grow from here.

For consumer markets: Direct impact: zero. What will actually reshape them is the next generation of closed-source products fed by these open models — cheaper, more Chinese-fluent local assistants.

BZH
K2 HorizonIFMDeepSeek·

K2 Horizon 把 375B 模型全套开源 — 大厂「伪开源」遇到对手了

上周,30 人不到的 IFM 团队一次性开源了 6 个模型(0.9B 到 375B)——这是 2025 年开源大模型最彻底的一次发布,也是对 DeepSeek、Llama「只开权重不开数据」做法最直接的挑战。

这是什么

IFM(Institute of Foundation Models)是加州一家专注「开放、独立」前沿模型的小型研究机构。K2 Horizon 是它一次性放出的模型矩阵:6 个模型覆盖手机端(0.9B)到 375B 旗舰。重要的是,它不只是把模型文件丢到 Hugging Face 上让你下载——训练数据配方、训练代码、中间权重 checkpoint(训练中途保存的快照)、每一步的训练日志和评测结果,全部公开。

对比一下:DeepSeek 公开权重但保留训练数据;Meta 的 Llama 设了商业使用条款;K2 Horizon 更接近学术界「完全可复现」的标准。

行业怎么看

支持者说这是对大厂「伪开源」的正面回应。当模型能力越强,闭源公司对开发者的控制力也越强——一个真正开放的模型,至少给学术界和中小企业留了一条退路。

反对意见同样尖锐。第一,375B 模型的推理和微调成本,普通团队根本承担不起,「全开源」对真实可用性影响有限。第二,Stable Diffusion 的故事已经上演过:模型红了一阵,后续维护和迭代跟不上,最终被闭源服务反超。第三,IFM 不到 30 人,融资和算力来源都没说清,能不能撑过下一轮训练是现实问题。

对普通人的影响

对企业 IT:短期不会有采购价值,但作为技术储备值得关注——尤其是 0.9B 这种小模型,未来可能跑在你公司的边缘设备上。

对个人职场:不必现在去下载。先记下这个名字;如果未来一年你听到越来越多公司在讲「私有部署」,背后的技术选择很可能从这里长出来。

对消费市场:直接感受为零。真正改变消费体验的,是被这些开源模型养出来的下一代闭源产品——更便宜、更懂中文的本地助手。