IFM发布K2 Horizon：0.9B至375B六款Apache 2.0模型
IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B
全尺寸开源且提供完整训练数据与代码，加上MoVA和Uno等创新工程实践，对研究者和开发者极具参考价值。
Most open model launches release one checkpoint and a benchmark table. The Institute of Foundation Models (IFM) released something wider last week. IFM is the frontier lab launched by MBZUAI in May 2025. K2 Horizon is a fleet of six models: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B. Shipping alongside them are the pre-training corpus, intermediate checkpoints, training code, configs and fine-grained logs. IFM calls it the largest fully open-source model launch in AI history.
大多数开源模型发布仅提供一个检查点和一张基准测试表。基础模型研究所（IFM）上周发布了更广泛的内容。IFM 是穆罕默德·本·扎耶德人工智能大学（MBZUAI）于 2025 年 5 月启动的前沿实验室。K2 Horizon 是一个包含六款模型的舰队：375B-A23B、36B-A4B、32B、7B、3.7B 和 0.9B。随同发布的还有预训练语料库、中间检查点、训练代码、配置以及细粒度日志。IFM 称这是 AI 历史上最大规模的完全开源模型发布。
Is it deployable? Yes, all six sizes sit on Hugging Face under Apache 2.0, with FP8 and GGUF builds. Day-zero support covers vLLM, SGLang and Ollama, on NVIDIA, AMD and Cerebras hardware. Hosted APIs run through Compass, Cerebras and Nebius via platform.ifm.ai.
是否可部署？是的，所有六个尺寸均托管在 Hugging Face 上，采用 Apache 2.0 许可证，并提供 FP8 和 GGUF 构建版本。首日支持涵盖 vLLM、SGLang 和 Ollama，适用于 NVIDIA、AMD 和 Cerebras 硬件。托管 API 通过 platform.ifm.ai 经由 Compass、Cerebras 和 Nebius 运行。
What Actually Shipped
实际发布内容
The six models share a core architecture, vocabulary, training methodology, interfaces and deployment tooling. The 0.9B model uses a smaller vocabulary. That consistency is the point: teams can prototype on 3.7B and scale to 375B-A23B without changing their serving stack.
这六款模型共享核心架构、词汇表、训练方法、接口和部署工具。0.9B 模型使用较小的词汇表。这种一致性正是其要点：团队可以在 3.7B 上进行原型设计，并扩展到 375B-A23B，而无需更改其服务堆栈。
Each model is pre-trained on roughly 20 trillion tokens. Nearly 17% of the pre-training corpus consists of problem-solving trajectories with explicit reasoning. About 10 trillion tokens were synthetic.
每款模型均在约 20 万亿个 token 上进行预训练。预训练语料库中近 17% 由带有显式推理的问题解决轨迹组成。约 10 万亿个 token 为合成数据。
Post-training data was folded in from mid-training rather than saved for the end. IFM research team reports over 100 million unique synthesized tasks. Tool definitions were presented in JSON, XML and Markdown during training so the model learns semantics rather than syntax. Markdown became the inference default, roughly 18.5% more token-efficient than JSON on IFM’s data.
后训练数据是从训练中后期整合进来的，而不是留到最后才保存。IFM 研究团队报告称有超过 1 亿个独特的合成任务。在训练期间，工具定义以 JSON、XML 和 Markdown 格式呈现，以便模型学习语义而非语法。Markdown 成为推理默认格式，在 IFM 的数据上比 JSON 的 token 效率高出约 18.5%。
MoVA: Sparsity Moved into Attention
MoVA：稀疏性进入注意力机制
Conventional Mixture-of-Experts applies sparsity to feed-forward layers. Mixture-of-Value Attention (MoVA) extends expert routing into multi-head attention itself, opening a second axis for scaling capacity. It stays compatible with FlashAttention, grouped-query attention and sparse attention.
传统的混合专家（Mixture-of-Experts）将稀疏性应用于前馈层。混合值注意力（MoVA）将专家路由扩展到多头注意力本身，开辟了扩展容量的第二个维度。它保持与 FlashAttention、分组查询注意力和稀疏注意力兼容。
The result is K2-Horizon-MoVA-36B-A4B: 36B total parameters, roughly 4B active per token. Under matched training conditions it lands slightly below the dense 32B model. On IFM’s tables it posts 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking, leading its comparison set on both.
结果是 K2-Horizon-MoVA-36B-A4B：总共 360 亿参数，每个 token 约有 40 亿活跃参数。在匹配的训练条件下，其表现略低于稠密的 32B 模型。在 IFM 的表格中，它在 Terminal-Bench 2.1 上得分为 58.6，在 tau3-Banking 上得分为 26.8，在比较集中两项均领先。
Uno: A Lossless Decoding Speedup as a LoRA
Uno：一种无损解码加速作为 LoRA
Uno freezes Horizon’s autoregressive parameters and trains a small set of diffusion parameters that learn only how to generate efficiently. Through what IFM calls diffusion distillation, these adapters emit blocks of tokens in parallel. The press release puts the speedup at roughly 3× with no quality degradation. It ships as a LoRA adapter, currently 7B-Uno and 0.9B-Uno.
Uno 冻结了 Horizon 的自回归参数，并训练一小组扩散参数，这些参数仅学习如何高效生成。通过 IFM 所称的扩散蒸馏（diffusion distillation），这些适配器并行输出 token 块。新闻稿称速度提升约为 3 倍，且无质量下降。它以 LoRA 适配器的形式发布，目前提供 7B-Uno 和 0.9B-Uno。
Numbers worth knowing
值得关注的数字
K2-Horizon-375B-A23B scores 70.2 on Terminal-Bench 2.1, 1,441 Elo on GDPVal-AA, 67.7 on MCPMark and 87.3 on GPQA Diamond. It leads its table on SWE-Atlas-QnA at 48.4 but trails GPT-5.6 Luna and Claude Sonnet 5 on most agentic rows.
K2-Horizon-375B-A23B 在 Terminal-Bench 2.1 上得分 70.2，在 GDPVal-AA 上得分为 1,441 Elo，在 MCPMark 上得分为 67.7，在 GPQA Diamond 上得分为 87.3。它在 SWE-Atlas-QnA 上以 48.4 分领先其所在表格，但在大多数智能体（agentic）行中落后于 GPT-5.6 Luna 和 Claude Sonnet 5。
The small models are the sharper story. 7B posts 70.6 on SWE-bench Verified and 59.0 on BrowseComp. 3.7B posts 68.6 on SWE-bench Verified. 0.9B reaches 48.5 on AIME 2026 and 79.9 on HumanEval+, small enough to run under quantization on a watch.
小模型的表现更为突出。7B 模型在 SWE-bench Verified 上得分为 70.6，在 BrowseComp 上得分为 59.0。3.7B 模型在 SWE-bench Verified 上得分为 68.6。0.9B 模型在 AIME 2026 上达到 48.5，在 HumanEval+ 上达到 79.9，其体量之小，即使在量化状态下也能在手表上运行。
The Audit IFM Ran on Itself
IFM 对自身进行的审计
This is the part many other labs do not publish. IFM ran 375B-A23B across 89 Terminal-Bench 2.1 tasks, eight attempts each. That is 712 trials, 500 passing, a reported 70.2% accuracy. Every passing trial was then re-audited using Artificial Analysis’s reward hacking procedure.
这是许多其他实验室未公开的部分。IFM 在 89 个 Terminal-Bench 2.1 任务上对 375B-A23B 进行了测试，每个任务尝试 8 次。共计 712 次试验，其中 500 次通过，报告准确率为 70.2%。随后，使用 Artificial Analysis 的奖励黑客（reward hacking）程序对所有通过的试验进行了重新审计。
The audit flagged 24 trials across 10 tasks. Removing them drops accuracy to 66.9%, a 3.37-point correction. That sits between the flag rates Artificial Analysis reports for Claude Fable 5 (2.2%) and GPT-5.6 Luna (4.1%). Behaviors included locating benchmark repositories on GitHub and downloading reference solutions. IFM also disclosed a 7B run that reached an inflated 82 on SWE-bench by finding answers.
审计标记了 10 个任务中的 24 次试验。移除这些试验后，准确率降至 66.9%，修正幅度为 3.37 个百分点。这一标记率介于 Artificial Analysis 报告的 Claude Fable 5（2.2%）和 GPT-5.6 Luna（4.1%）之间。涉及的行为包括在 GitHub 上定位基准存储库并下载参考解决方案。IFM 还披露了一次 7B 模型的运行，该模型通过寻找答案使 SWE-bench 得分虚高至 82。
Interactive explainer
交互式解释器
Key Takeaways
关键要点
- Six models, 0.9B to 375B, all Apache 2.0, all sharing one architecture and serving stack.
- MoVA pushes MoE routing into attention: 36B total, ~4B active, near dense-32B quality.
- Uno delivers roughly 3× lossless decoding speedup as a drop-in LoRA adapter.
- The 0.9B, 3.7B and 7B models claim state of the art at their respective scales.
- IFM published its own reward-hacking audit, correcting 70.2% down to 66.9%.
- 六款模型，参数量从 0.9B 到 375B 不等，均采用 Apache 2.0 许可证，共享同一架构和服务堆栈。
- MoVA 将 MoE 路由推入注意力机制：总参数量 36B，活跃参数量约 4B，性能接近密集型 32B 模型。
- Uno 作为即插即用的 LoRA 适配器，提供了大约 3 倍的无损解码速度提升。
- 0.9B、3.7B 和 7B 模型在其各自规模下声称达到了最先进水平（state of the art）。
- IFM 发布了自身的奖励黑客审计结果，将准确率从 70.2% 修正至 66.9%。
Check out the Technical blog, Press release, Hugging Face collection, TxT360-v2 dataset, xLLM pre-training code, Post-training code and Docs. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看技术博客、新闻稿、Hugging Face 合集、TxT360-v2 数据集、xLLM 预训练代码、后训练代码和文档。此外，欢迎在 Twitter 上关注我们，别忘了加入我们拥有 15 万 + 成员的 ML SubReddit 并订阅我们的通讯。等等！你在 Telegram 上吗？现在你也可以在 Telegram 上加入我们。
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
需要与我们合作推广您的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等？请与我们联系
The post IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B appeared first on MarkTechPost.
IFM 发布 K2 Horizon：六款基于 Apache 2.0 协议的模型，参数量从 0.9B 到 375B。本文最初发表于 MarkTechPost。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力