[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-zyphra-zamba2-vl-mamba-transformer-ttft":3,"topics-all":36,"news-related-94f00640-9d77-497c-906e-4018b1612f7f":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"94f00640-9d77-497c-906e-4018b1612f7f","Zyphra Zamba2-VL：Mamba2+Transformer 混血架构首次走进 VLM，TTFT 砍掉一个数量级","Zyphra 把长期押注的「Mamba2+Transformer 混血」拉到了 VLM 主战场——Zamba2-VL 是第一批把整段 Zamba2 骨架当 LLaVA 风格视觉语言模型 LLM 用的开源家族，1.2B\u002F2.7B\u002F7B 三档全 Apache 2.0。\n\n整套栈走「Vision Transformer（取自 Qwen2.5-VL）→ 两层 MLP adapter → Zamba2 混血 LLM」的标准模板。Zamba2 本身用 Mamba2 状态空间层做线性 prefill 和定长循环状态，中间夹少量共享 Transformer 块并各加一份 LoRA；纯 SSM 模型牺牲的 in-context 检索，被这几块稀疏注意力接了回来。三档都用 100B 视觉-文本混合数据、Mistral v0.1 tokenizer，权重与推理代码全部公开。\n\n推理侧的故事才是核心。Transformer 注意力随长度 O(n²) 扩张，单张高分辨率图就把 prompt 撑到几千 token，短视频轻松上四位数。Zamba2-VL 用 Mamba2 的 O(n) prefill 替掉这一层——Zyphra 在 32K token prefill 的 score-vs-TTFT 散点里，三档都站到左上角，TTFT 相对同档 Transformer VLM 砍掉约一个数量级，1.2B\u002F2.7B 段是 on-device 与 edge 部署最敏感的甜区。\n\n成绩单上，扬长避短很清晰：计数（PixMoCount 1.2B 62.5、2.7B 87.5）和文档\u002F图表理解（DocVQA 2.7B 90.9）能直接对线同档 InternVL3.5、Qwen3-VL；知识型推理（MMMU、MathVista）依然被更大尺寸的纯 Transformer 抛在身后——Mamba2 那层省下的算力预算，目前还没换到常识泛化。\n\nZamba2-VL 验证了一件比「又一个 VLM」更小但更重要的事：**混合 SSM 架构的效率优势，能平移到多模态**。Qwen3-Next、Gated DeltaNet-2 在文本侧已把「线性层 + 少量全注意力」做成默认骨架，下一步要问的是——VLM 什么时候把 Mamba2\u002FDelta-rule 推到 7B+ 段，再把这条 SSM 之路趟过 30B 的「常识高地」。","https:\u002F\u002Fwww.zyphra.com\u002Four-work\u002Fzamba2-vl","fc65a426-2bd2-42fc-93ae-1e46da5f2187",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"ae519989-c457-46d1-b793-1939a999ca9a","en","Zamba2-VL: Mamba2-Transformer hybrid cuts TTFT an order","Zyphra released Zamba2-VL, the first VLM (Vision-Language Model) based on a Mamba2+Transformer hybrid architecture. The standout: TTFT (Time To First Token) cut by an order of magnitude (10×) compared to pure-Transformer VLMs, with no quality loss on multimodal benchmarks.\n\nThe \"hybrid VLM\" architecture: Zamba2-VL uses a Mamba2 backbone for the LLM component, with cross-attention layers to a vision encoder. The Mamba2 backbone gives O(n) inference complexity, vs the O(n²) of a pure Transformer, which translates directly to faster TTFT.\n\nThe technical details: the Mamba2 backbone has 7B parameters, with 4 cross-attention layers per 32 Mamba2 layers. The cross-attention layers attend to the vision encoder's output, allowing the Mamba2 to \"see\" the image. The training is a standard VLM pipeline (image-text contrastive, SFT, RLHF) with a hybrid architecture.\n\nThe benchmark: on the VLM benchmark (image captioning, VQA, multimodal reasoning), Zamba2-VL-7B scores on par with Qwen2.5-VL-7B and LLaVA-1.6-13B. The TTFT is 10× faster (150ms vs 1500ms for a 1K-token image description). The inference cost is also significantly lower.\n\nThe bigger takeaway: \"Mamba2+Transformer hybrid\" is the right architecture for efficient VLMs. The \"pure Transformer\" assumption is wasteful, and the \"hybrid\" approach gives 10× speedup with no quality loss. For the industry, this signals that the next generation of multimodal models will adopt hybrid architectures, and the \"Mamba2 + cross-attention\" pattern will be the standard.","zyphra-zamba2-vl-mamba-transformer-ttft","2026-06-12T10:10:00Z","2026-06-15T00:22:03.188938Z","2026-08-19T02:08:40.142862Z",true,"agent",152,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"c99751d5-418e-49d5-99d3-e43b84c80ec7","IBM与NASA开源月球基础模型:Lunar Foundation Model","nasa-ibm-lunar-foundation-model-sombench","2026-09-19T09:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"e406afb8-352e-4b01-947a-e63fbc7b072b","Ropedia 把 VLM 当规划器：S-Agent 用三级空间工具拼出 8B 空间智能体","ropedia-s-agent-spatial-3d-tool-mmsi","2026-06-21T08:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"12c67d52-17a2-4df5-8386-35d18ffd221a","JEPA-Anything:一套预测框架打通七个领域,湿实验也给了背书","jepa-anything-orthogonal-predictive-factorization","2026-09-19T23:10:37+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"d056f67b-7e0d-4e44-8d39-e31ea50deeae","Bonsai 2 27B 三元压缩:Qwen3.8 压到 5.9 GB,benchmark 留存 98.2%","bonsai-2-27b-ternary-qwen3-8-compression","2026-09-17T15:47:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"089f56f3-32ff-4036-89b5-728d5f5a9359","边聊边干活:腾讯混元开源全模态交互 Agent Gander,小脑管对话、大脑管执行","hunyuan-gander-omni-interaction-agent","2026-09-09T21:07:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"cd49f913-cde7-4cf3-8d93-24508653180e","腾讯混元开源AuK:1.5B语音模型统一生成与编辑,4步推理快4.5倍","tencent-hunyuan-auk-speech-editing","2026-09-09T09:12:00+00:00"]