[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kimi-k3-deepseek-v4-native-multimodal-divergence":3,"news-related-2d29aa3d-317c-4126-a6f7-2c9c2c3b6f93":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"2d29aa3d-317c-4126-a6f7-2c9c2c3b6f93","Kimi K3与DeepSeek V4之间,隔着原生多模态的时间差","2026 年 8 月,国产大模型头部出现明显技术路线分歧:月之暗面 Kimi K3 走 2.8 万亿参数 MoE + 原生视觉融合路线,DeepSeek V4-Flash 仅 2840 亿参数坚持纯文本后训练路线。两条路在 Coding 与 Agent 榜单上短兵相接,但对原生多模态的态度决定了它们眼下的形态和长期演化的节奏。本文逐条拆解它们为何同台而不同路。","## Coding 之外,关于原生多模态的争论从未停止\n\n过去一年,Coding 与 Agent 能力不断改写大模型排名,也是当前 AI 最快兑现商业价值的场景之一;但 Coding 之外,另一个更长期的命题正在被拉回到桌面中央:通用大模型要不要长出「眼睛」?\n\n刚刚过去的 7 月,月之暗面发布 Kimi K3,总参数 2.8 万亿、支持 100 万 token 上下文,延续 K2.5 的 early fusion(早期融合)路线,在预训练阶段就把图文按固定比例混合,同时从零训练了约 4 亿参数的视觉编码器 MoonViT-V2,不再依赖 SigLIP 初始化权重。K3 在 Arena Frontend Code 榜单上曾以 1679 分登顶,浏览器开发平台 Puter 在测试页中故意埋了 5 处视觉偏差,K3 对照目标页与运行页截图,全部找出且零误报。月之暗面把这种「写完代码→看页面→再改」的循环命名为 *vision in the loop*:模型自己看自己写出来的东西。\n\n而同期走向 GA 的 DeepSeek 走了另一种答案。DeepSeek V4-Flash 正式版总参数 2840 亿、每个 token 激活 130 亿,分别约为 K3 的十分之一和八分之一。它不调整架构、也不扩张参数规模,而是把更新集中在后训练。结果是:在 DeepSeek 公布的多项 Coding 与 Agent 评测中,正式版大幅超过预览版,甚至超过此前的 V4-Pro 预览版;Arena WebDev 公开得分也一度升至 1577,接近 GLM-5.2。\n\n一句话总结:同样是在 Coding 这一张牌桌上,K3 选择为大模型额外接入视觉感知,V4-Flash 选择把 Coding 分数硬打到极致。\n\n## 为什么 Agent 必须「看见」\n\n一位多模态研究员把核心论点讲得很直白:「长链任务如果只通过代码层面的反馈,误差可能会不断累积,最终效果会非常差。视觉是一种更准确的反馈,也更贴近用户意图。」\n\n纯文本模型本身仍然「看不见」。实际系统可以调 OCR、独立 VLM 或 MCP 接入的视觉工具——把图片转成文字、标签、坐标或结构化字段,再交给主模型推理。这条「外挂」路线工程上完全可行,但原生多模态派认为它有两个天花板:一是画面先被压缩成文字,带宽变窄;二是后训练时要让视觉反馈纳入同一条训练轨迹比较难。\n\n杨立昆与 Ilya Sutskever 的分歧也适用于此:杨立昆认为「绝大多数人类知识,并没有以文本形式表达出来」,模型要真正理解物体、空间和行动,最终仍要从图像、视频和现实交互中学习;Ilya Sutskever 则坚持「文本是世界的一种投射」,只要模型持续吃文本,就能获得对世界的理解。两条路径都不否认多模态的价值,分歧只在时机与代价。\n\n## 多模态的代价:不是 1+1,而是大于 2\n\n把视觉接进 LLM,看起来只是新增一个输入接口,实际却要面对结构性冲突。图像与文字的数据结构、信息密度和学习速度都不同;它们共享同一套主干参数时,会和文本、代码、数学、推理争夺容量,不同训练目标也会相互干扰。\n\nKimi K2.5 的技术报告给出一个直观的消融证据:如果把视觉数据放到训练中后期才加入,模型的文本能力会先下降,再缓慢恢复。换句话说,「1+1 大于 2」是真实存在的成本——vision-in-the-loop 做得越好,语言\u002FCoding 的稳定性越需要重新平衡。\n\n另两个隐性成本:算力与数据。苹果 MM1 论文指出,视觉编码器、图像分辨率、视觉 token 数量都会显著影响效果;更高的分辨率与更多视觉 token 通常意味着更高的训练与推理开销。而对那些只需读几个字段的任务,让通用模型反复查看整张图片,未必比 OCR 或专业模型更划算。\n\n## 国产头部已经分裂成两条路线\n\n放眼 2026 年 8 月的国产通用基座阵营,路线选择已经肉眼可见:\n\n- **原生多模态派**:月之暗面(Kimi K3,2.8T\u002F104B)、阿里(Qwen3.8-Max,2.4T\u002F95B,同时承载原生视觉、Coding 与 Cowork 能力)、字节(Doubao-Seed-2.1,视觉理解和多模态输入作为主要功能)——共同特征是「水桶模型」,在大参数规模下原生融合视觉;\n- **纯文本后训练派**:DeepSeek(V4-Flash 284B\u002F13B)、智谱(下一代通用基座被业内预期将进一步向多模态靠拢)、腾讯混元——基座仍以文本输入为主,先把 Coding\u002FAgent 分数拉满。\n\nDeepSeek 创始人梁文锋那句「把 AI 训练做好,并不需要世界模型,甚至不用多模态」,常被引用为「不看好多模态」的证据;他后半句补得也很直接——「多模态最终还是要做的。」这两句话并不矛盾,真正的分歧在于:Coding 还在快速迭代的阶段,要不要同步花资源训练视觉?\n\n## 写作这篇文章的意义:两个时钟在赛跑\n\n如果只看 Coding 与 Agent 的短期排名,这是一场几个月甚至几周就会重新洗牌的短跑;但如果把视野拉到世界模型这个维度,通用大模型从纯文本走向真正的视觉理解,是一场以年为单位的马拉松。\n\nK3 在跑前一个时钟——它用 2.8 万亿参数、native 视觉、100 万上下文把 Arena Frontend Code 的上限往上顶了一截;V4-Flash 也在跑前一个时钟,它用 1\u002F10 的参数规模、后训练蒸馏,在 Arena WebDev 反超了一众预览版。这两条路短期都有效,但分歧点决定了它们最终能抵达的地方不一样:\n\n- 当 vision in the loop 成为 Coding Agent 的标配反馈通道时,纯文本 + 后训练的路线会撞上「看得见看不见」的天花板;\n- 而当下一次多模态 benchmark 重新定义 SOTA 时,提前把视觉烧进主干的模型会立刻处在有利位置。\n\n## 所以呢\n\n对中国大模型而言,这不是关于 AGI 长期路线的分歧,更像是对发展时机和代价的不同押注。眼下商业化压力把 Coding 顶到牌桌门槛——Coding 冲不上去,可能就没有下一轮。但在 Coding 战事胶着的间隙,各家真正在投注的下一个筹码,是愿意提前为视觉留出多少模型容量、数据与算力。\n\n可以确定的一点是:2026 下半年,围绕「原生多模态」的讨论还会持续升温。无论是 Puter 那 5 处被找出的视觉偏差、还是 V4-Flash 用 2840 亿参数在 WebDev 拿到 1577 分,都是这条路线之争的具体注脚。\n\n参考资料:\n- 36氪《Kimi K3与DeepSeek V4之间,隔着原生多模态的时间差》(李炤锋,2026-08-04):https:\u002F\u002F36kr.com\u002Fp\u002F3924826666301831\n- Kimi K3 发布与 Arena Frontend Code 榜单表现(同上来源)  \n- DeepSeek V4-Flash 正式版参数与 Arena WebDev 得分(同上来源)","https:\u002F\u002F36kr.com\u002Fp\u002F3924826666301831","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"b52db7e9-7c58-42c3-9536-5132cb2f8f72","deepseek",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"20001057-aff1-4768-8976-a7b16037af12","en","Beyond Coding: Kimi K3 and DeepSeek V4 Diverge on Native Multimodality","In August 2026, China's top foundation-model houses took visibly different routes: Moonshot's Kimi K3 committed to a 2.8T-parameter MoE with native visual fusion, while DeepSeek's V4-Flash stayed at 284B parameters and a pure-text post-training regime. The two paths collide on Coding and Agent leaderboards, but the attitude toward native multimodality decides both their current shape and their long-term evolution. This piece walks through why they share a leaderboard but walk different roads.","## Beyond Coding: The Native-Multimodal Debate Is Still Alive\n\nOver the past year, Coding and Agent capabilities have rewritten the LLM leaderboards — they are also the fastest path to commercial value for AI. But beside that race, another, longer-running question is back on the table: should general-purpose LLMs be given *eyes*?\n\nJuly 2026 saw Moonshot release Kimi K3 — a 2.8-trillion-parameter MoE with a 1-million-token context window, continuing the early-fusion recipe from K2.5. Visual and text tokens are mixed at a fixed ratio from the very start of pre-training, and the visual encoder — MoonViT-V2 (~400M parameters) — is trained from scratch rather than being initialized from SigLIP. On the Arena Frontend Code leaderboard, K3 hit 1679 points at one point. The browser-platform company Puter deliberately planted 5 visual discrepancies in a test page; K3 compared the target page with the rendered output and found all five, with zero false positives. Moonshot calls this loop \"vision in the loop\": the model writes code, inspects the result, and adjusts.\n\nAround the same time, DeepSeek took a different route. The GA version of DeepSeek V4-Flash has 284B total parameters and activates only 13B per token — roughly one-tenth and one-eighth of K3's figures, respectively. No architecture change, no parameter inflation — just a sharp post-training push. The payoff shows up on leaderboards: V4-Flash (formal) substantially beat its own preview and even the V4-Pro preview across DeepSeek's reported Coding and Agent evaluations, and its Arena WebDev public score briefly climbed to 1577, putting it within reach of GLM-5.2.\n\nThe short version: on the same Coding table, K3 bets on giving the model real visual perception, while V4-Flash bets on grinding Coding scores to their absolute ceiling with much less compute.\n\n## Why Agents Must *See*\n\nOne multimodal researcher puts the argument simply: \"On long-horizon tasks, errors compound when feedback only comes from the code path. Vision is a more accurate signal and a closer match for user intent.\"\n\nPure-text LLMs are still \"blind.\" Real systems can wire in OCR, an independent VLM, or vision tools over MCP — converting images into text, tags, coordinates, or structured fields before handing them to the main model. That \"add-on\" path is perfectly workable in production, but the native-multimodal camp argues it has two ceilings: the image is compressed into text first, so the bandwidth narrows; and pulling visual feedback into a single training trajectory during post-training is much harder.\n\nThe same disagreement shows up at the field level. Yann LeCun's view: \"the vast majority of human knowledge isn't expressed as text.\" Image, video, and real-world interaction remain the source from which any model that wants to understand objects, space, and action eventually has to learn. Ilya Sutskever's opposing take: \"text is a projection of the world,\" and a model that keeps learning from text can still build world understanding. Both sides agree on the long-run value of multimodality. The disagreement is about *when* and at *what cost*.\n\n## The Cost of Multimodality: Not 1+1, but >2\n\nAdding vision to an LLM looks like adding another input port. Underneath, it is a structural collision. Images and text have different data structures, information densities, and learning rates. Sharing one backbone forces them to fight for capacity against code, math, reasoning and each other, and training objectives interfere.\n\nKimi K2.5's technical report gives a concrete ablation: when visual data is added late in training, the model's text capability drops first and only slowly recovers. In other words, \"1+1 > 2\" is a real cost — the better vision-in-the-loop works, the more carefully text\u002FCoding stability has to be rebalanced.\n\nTwo hidden costs compound that: compute and data. Apple's MM1 paper showed that visual encoder choice, image resolution, and the visual-token count all materially affect results; higher resolutions and more visual tokens mean higher training and inference cost. For tasks that only need to read a few fields, asking a general model to keep scanning the whole image is not obviously cheaper than running OCR or a specialist model.\n\n## China's Top Tier Has Already Split\n\nBy August 2026, the Chinese general-purpose foundation-model camp splits into two visible tracks:\n\n- **Native-multimodal camp**: Moonshot (Kimi K3, 2.8T \u002F 104B activated), Alibaba (Qwen3.8-Max, 2.4T \u002F 95B activated, combining native vision, Coding and Cowork), ByteDance (Doubao-Seed-2.1, with visual understanding and multimodal input as headline features) — all \"bucket models\" that natively fuse vision into very large parameter budgets.\n- **Pure-text post-training camp**: DeepSeek (V4-Flash at 284B \u002F 13B), Zhipu (next-generation general base model expected to move further toward multimodality), Tencent HunYuan — text-in, text-out, with Coding\u002FAgent scores optimized first.\n\nDeepSeek founder Liang Wenfeng once said: \"To train AI well, you don't need a world model — you don't even need multimodality.\" That line is often quoted as 'DeepSeek is bearish on multimodality.' It is less often quoted with the second half: \"But multimodality is something we'll eventually have to do.\" Both halves are true. The real split is this: while Coding is still iterating fast, how much model capacity, data, and compute do you spend on training vision at the same time?\n\n## Two Clocks Are Racing\n\nIn the short run, Coding and Agent rankings are reshuffled every few weeks or months — that is one clock. In the long run, going from text-only LLMs to genuinely vision-grounded understanding is a multi-year marathon — that is another.\n\nK3 is racing the first clock — using 2.8T parameters, native vision, and a 1M context to push Arena Frontend Code's ceiling higher. V4-Flash is also racing the first clock — using one-tenth the parameter count and post-training distillation to overtake a stack of previews on Arena WebDev. Both strategies work right now. But the divergence shapes how far each can eventually go:\n\n- When vision-in-the-loop becomes the default feedback channel for Coding Agents, the pure-text + post-training line will hit a \"see \u002F don't see\" ceiling.\n- When the next multimodal benchmark redefines SOTA, the models that baked vision into the backbone earlier will start from a structurally stronger position.\n\n## So What\n\nThis is not a disagreement about the long-run AGI path for Chinese foundation models. It is a different bet on timing and cost. Right now, commercial pressure keeps Coding as the price of admission — no Coding breakthrough, no next round. But in the gaps between Coding battles, the real next chip each house is buying is *how much model capacity, data, and compute it is willing to set aside for vision, ahead of time*.\n\nOne thing is certain: through the second half of 2026, the debate over \"native multimodality\" is going to keep heating up. Puter's 5 caught visual discrepancies and V4-Flash reaching 1577 on WebDev with just 284B parameters are concrete footnotes in that ongoing argument.\n\nReferences:\n- 36Kr, \"Kimi K3 vs DeepSeek V4: the native-multimodal time gap\" (Li Zhaofeng, 2026-08-04): https:\u002F\u002F36kr.com\u002Fp\u002F3924826666301831\n- Kimi K3 release & Arena Frontend Code performance (same source)\n- DeepSeek V4-Flash GA parameters and Arena WebDev score (same source)","kimi-k3-deepseek-v4-native-multimodal-divergence","2026-08-04T08:02:10Z","2026-08-04T08:02:36.142217Z","2026-08-04T08:02:36.142233Z",true,"agent",233,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"d7b6d14d-7257-4794-b92f-31956bbc7eae","原生多模态 vs 后训练加压:国产头部基模两条路线的工程账","native-multimodal-vs-posttraining-2026","2026-08-05T00:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"31f3215c-0892-419d-a610-fe815cc60bbe","GPT-5.6 降价 80% 把竞争拉进「同等智能成本」：DeepSeek V4 Flash 接招，国产模型卡出双线赛道","gpt-5-6-luna-price-cut-equal-intelligence-cost","2026-08-12T03:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"34edaffc-6b5c-4df1-9e2f-d864cada6063","Gemini 走进 K-12 课堂：Google 把「上下文」塞进每个作业","gemini-classroom-k12-contextualized-prompts","2026-08-07T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"c5304aea-6b80-4fbb-b2ac-f07a7a6d7713","DeepSeek V4 Flash 上午短暂\"翻车\"：国产开源模型的容量大考","deepseek-v4-flash-api-capacity-incident","2026-08-04T10:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"4b693fb4-541f-47ed-8923-6e280cec965f","大模型的“记忆”还没过视觉这一关：MEMLENS 把长上下文的短板测出来了","memlens-multimodal-long-term-memory","2026-08-03T02:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"96f87320-5237-44cc-9919-dcee43a6ef80","DeepSeek-V4-Flash-0731 正式版 API 公测:只换后训练不换权重,开源 LLM 进入轻量迭代节奏","deepseek-v4-flash-0731-post-training-ga","2026-07-31T10:00:00+00:00"]