[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-native-multimodal-vs-posttraining-2026":3,"news-related-d7b6d14d-7257-4794-b92f-31956bbc7eae":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"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","围绕月之暗面 Kimi K3 与 DeepSeek V4-Flash 的同期发布,本文剖析国产头部基模厂商在 Agent 时代对'是否同步训练视觉'的不同答案:K3 走 2.8 万亿参数 MoE + 原生多模态路线,Arena WebDev 跑到 1679 分;DeepSeek V4-Flash 用 2840 亿参数 + 后训练加压,也能在 WebDev 上拿到 1577 分。背后的资源取舍与时机判断,正在重写国产基础模型的训练哲学。","2026 年 7 月,月之暗面 Kimi K3 与 DeepSeek V4-Flash 几乎前后脚发布。两款模型在 Coding 与 Agent 榜单上的位次咬得很紧,但走的却是完全不同的两条技术路线。\n\n**K3 是把视觉一起拉进主模型;V4-Flash 是把参数规模按下去,只靠后训练。**\n\nK3 总参数 2.8 万亿,每个 token 激活约 1040 亿,延续了 K2.5 的 early fusion 路线——在约 15 万亿混合图文 token 的联合预训练中,始终按固定比例混合文本与视觉数据,并从零训练了一个 4 亿参数的视觉编码器 MoonViT-V2。K3 发布后一度以 1679 分登顶 Arena Frontend Code,前面只剩 Claude Fable 5、GPT-5.6 Sol 等少数对手(Arena WebDev 上 K3 与 V4-Flash 也都跑进了 Top 区间,前者 1679 分,后者 1577 分)。\n\nDeepSeek V4-Flash 选择不去掺和视觉。它的总参数只有 2840 亿,每个 token 激活约 130 亿,大约只有 K3 的十分之一和八分之一。模型没有调整架构和参数规模,而是把更新集中在后训练——结果,在 DeepSeek 自己公布的 Coding 和 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。\n\n紧接着阿里发布的 Qwen3.8-Max,又选择了与 K3 相近的\"大而全\"方向:总参数 2.4 万亿,每个 token 激活约 950 亿,同时承载原生视觉、Coding 与 Cowork 能力。三家头部基模,排成两种路线:Kimi 与 Qwen 押宝\"超大参数 + 原生多模态\",DeepSeek 押宝\"小模型 + 后训练加压\"。\n\n## Agent 为什么需要眼睛\n\n视觉对 Agent 越来越关键。OpenAI 1 月发布的企业使用报告显示,在研发岗位最常使用的三类 ChatGPT 工具中,图片上传排第三位,仅次于搜索和数据分析。当 Agent 开始生成网页、操作软件并检查运行结果,视觉就成了模型检查工作、发现错误和调整行动的反馈。\n\nKimi 方面把这种\"在代码与截图之间反复迭代\"的方式称为 \"vision in the loop\"——模型写完代码后查看页面,再根据视觉结果继续调整。浏览器开发平台 Puter 曾在测试网页中制造 5 处视觉偏差,K3 对照目标页与运行页截图全部找出,没有误报。\n\n相比之下,纯文本模型只能靠 OCR 或外挂视觉语言模型先把图片转成文字、标签、坐标或结构化字段。在一位多模态研究员看来,这种\"一个 LLM 是'瞎子',下游配一个能看清楚的小弟\"的模块化方案有边界——视觉输入与语言主干之间\"通信的通道会更宽\",不是先把画面压缩成文字再交给另一个模型。原生多模态模型还可以在视觉强化学习中重新读取自己生成的页面或图像,把视觉结果纳入同一条训练轨迹。\n\n但视觉是不是模型理解世界不可缺少的部分,业内始终没有共识。OpenAI 联合创始人、前首席科学家 Ilya Sutskever 曾把文本称为\"世界的一种投射\":人类已经把大量现实规律压缩进语言,模型持续学习文本,仍可能获得对世界的理解。Meta 首席 AI 科学家 Yann LeCun 几乎持相反判断:\"绝大多数人类知识,并没有以文本形式表达出来。\"语言是经过人类筛选和概括的信息,模型要理解物体、空间和行动,最终仍要从图像、视频和现实交互中学习。\n\n## 视觉的代价:1+1 显著大于 2\n\n把视觉嵌入主模型,代价并不只是\"接入一个输入\"。\n\n图像与文字的数据结构、信息密度和学习速度不同。当它们共同训练同一套主干参数时,视觉数据会与文本、代码、数学和推理争夺模型容量;不同训练目标也可能相互干扰。Kimi K2.5 技术报告在一项融合时机消融实验中记录过这种影响:如果在训练中后期才加入视觉数据,模型的文本能力会先下降,再逐渐恢复。\n\n\"任务越多,越难平衡。\"一位多模态研究员说,\"只做一个任务,可以一直往一个方向调;同时做两个任务(视觉 + 文本),要么模型更大一点,要么数据更多一点。\"\n\n苹果公司的 MM1 技术报告也提示过类似结论:视觉编码器、图像分辨率和视觉 token 数量都会影响模型效果;更高的分辨率和更多视觉 token 通常也带来更高的训练与推理开销。对于只需读取几个字段的任务,让通用模型反复查看整张图片,不一定比 OCR 或专业模型来得划算。\n\n## 时机分歧:把算力留给多模态还是 Coding\n\nDeepSeek 创始人梁文锋此前说过一句话:\"把 AI 训练做好,并不需要世界模型,甚至不用多模态。\"但他同时表示:\"多模态最终还是要做的。\"这两句话并不矛盾。各家公司对多模态的长期价值没有太大分歧,真正的分歧在于时机和代价。\n\n6 月底,智谱首席科学家唐杰在 X 平台向用户征集\"下一版 GLM 必须加入哪些新功能\",评论区反复出现的答案是\"视觉\"。从公开口径看,智谱与腾讯混元的下一代通用基座模型,都有可能进一步向原生多模态迈进。\n\n但 DeepSeek V4-Flash 已经先一步证明,在不加入视觉、也不大幅扩张参数规模的情况下,后训练仍能显著提升 Coding 等核心任务能力。这就让当下的资源分配问题变得相当具体:把训练视觉的算力和数据预算,提前投入,还是先在 Coding 与 Agent 上把分数卷起来?\n\n多位业内人士认为,在 Agent 任务链越来越长的阶段,Coding 才是上牌桌的门槛,Coding 冲不上去可能就没有未来的机会。换句话说,Coding 与 Agent 能力的变现速度是几个月甚至几周;而模型从\"看得见\"走到\"看得懂\"则是一个更漫长的进化周期。\n\n## 评论:两个时钟同时在走\n\n围绕多模态,国产头部基模正在打两种时间节奏的赛跑。\n\n把视觉提前塞进主模型,意味着\"未来看现在\"——赌的是 Agent 接管更多长链任务、网页和 GUI 成为模型主要工作面之后,视觉反馈会从加分项变成入场券。K3、Qwen3.8-Max 押的就是这条路。\n\n把视觉放一放、把参数规模按下去、把所有力气压在后训练,意味着\"先守住眼下\"——赌的是 Coding 与 Agent 榜单会持续滚动,排名比\"看得见\"更靠近现金流。DeepSeek V4-Flash 押的就是这条路。\n\n这两条路线不是非此即彼。Kimi 的策略早就在做后训练,DeepSeek 早晚也会做原生多模态。但在 2026 年夏天这个时间点,它们各自在训练成本与产品决策之间画了一条明显的线——这条线与其说由技术结论决定,不如说由公司对节奏的判断决定。至于这两条线最终会在哪一年交汇,留给下一个交付周期去验证。\n\n*来源说明:本文事实点(模型参数、benchmark 分数、发布时间线、人物引语、技术报告结论)均回溯到原报道 [36 氪《Kimi K3 与 DeepSeek V4 之间,隔着原生多模态的时间差》](https:\u002F\u002F36kr.com\u002Fp\u002F3924826666301831?from=newsforai-twoclock) 及 Kimi K2.5 技术报告。*","https:\u002F\u002F36kr.com\u002Fp\u002F3924826666301831?from=newsforai-twoclock","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[11,15,18,21,24],{"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",{"id":25,"name":26,"slug":26,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"55e3cea1-f7aa-4f0e-a79f-07f63d887220","en","Native versus post-trained multimodality: China's two routes","In July 2026, Moonshot's Kimi K3 and DeepSeek V4-Flash shipped almost back-to-back. They sit close on Coding and Agent leaderboards, yet they take opposite engineering routes. K3 follows a 2.8-trillion-parameter MoE plus native multimodal path that scores 1679 on Arena WebDev. V4-Flash runs on 284 billion parameters with heavy post-training and lands at 1577 on the same board. The resource trade-off and the timing call are quietly rewriting how Chinese frontier models get built.","In July 2026, Moonshot's Kimi K3 and DeepSeek V4-Flash shipped almost back-to-back. They sit close on Coding and Agent leaderboards — yet they take opposite engineering routes.\n\n**K3 pulls vision into the main model. V4-Flash keeps parameter scale small and leans on post-training.**\n\nK3 has 2.8 trillion total parameters and activates roughly 104 billion per token. It continues the early-fusion recipe from K2.5 — a roughly 15-trillion-token joint pretraining that mixes text and vision tokens at a fixed ratio, plus a 4-billion-parameter vision encoder, MoonViT-V2, trained from scratch. After release, K3 briefly topped the Arena Frontend Code list at 1679, behind only Claude Fable 5, GPT-5.6 Sol and a handful of others. On Arena WebDev, K3 and V4-Flash both entered the top band — 1679 for K3, 1577 for V4-Flash.\n\nDeepSeek V4-Flash chose to stay out of vision. Total parameters are 284 billion, with about 130 billion activated per token — roughly one tenth and one eighth of K3 respectively. The architecture and scale barely move; all the work goes into post-training. According to DeepSeek's own Coding and Agent evaluations, the release version beat the preview version by a wide margin and even surpassed the earlier V4-Pro preview.\n\nAlibaba's Qwen3.8-Max, released right after, chose a direction close to K3's \"big and broad\" play: 2.4 trillion total parameters, roughly 95 billion activated per token, carrying native vision, Coding and Cowork at the same time. Three frontier base-model makers now line up into two camps — Kimi and Qwen bet on \"very large parameters plus native multimodal,\" DeepSeek bets on \"small model plus post-training pressure.\"\n\n## Why Agents need eyes\n\nVision is becoming increasingly critical for Agents. An OpenAI enterprise-usage report from January 2026 noted that, among the three ChatGPT tools most often used in research roles, image upload ranks third — after search and data analysis. As Agents begin to generate web pages, operate software and inspect runtime results, vision turns into the feedback channel that lets the model check work, spot bugs and adjust action.\n\nKimi calls this \"code-then-screenshot, iterate\" loop \"vision in the loop\" — write code, look at the page, adjust based on what is on screen. The browser-development platform Puter planted five visual discrepancies in a test page; K3 compared the target and rendered screenshots, found all five, and reported no false positives.\n\nPlain text models, by contrast, can only rely on OCR or an external vision-language model to convert images into text, tags, coordinates or structured fields. A multimodal researcher describes that \"tooled-up\" arrangement as bounded: the LLM is the 'blind person,' and a smaller model downstream is the 'eyes.' The communication channel between vision input and the language backbone inside a native multimodal model is \"wider\" — it does not compress an image into text before handing it to a separate model. Native multimodal models can also run visual reinforcement learning on their own rendered pages or images and fold the visual result back into the same training trajectory.\n\nBut whether vision is essential to a model understanding the world remains contested. Ilya Sutskever, OpenAI co-founder and former chief scientist, has called text \"a projection of the world\": humans have already compressed much of reality into language, and a model that keeps learning text can still grasp the world. Meta chief AI scientist Yann LeCun argues the opposite: \"The vast majority of human knowledge is not expressed in textual form.\" Language is human-selected and summarized; if a model is to understand objects, space and action, it must ultimately learn from images, video and real-world interaction.\n\n## The cost of vision: 1 + 1 is meaningfully greater than 2\n\nFolding vision into the main model costs more than \"plugging in another input.\"\n\nImage and text differ in data structure, information density and learning speed. When they are trained against the same backbone parameters, vision data competes with text, code, math and reasoning for model capacity, and different training objectives can interfere. The Kimi K2.5 technical report recorded this in an ablation on fusion timing: if visual data is added late in pretraining, text capability first drops and then slowly recovers.\n\n\"The more tasks, the harder the balance,\" a multimodal researcher says. \"Doing one task, you can keep tuning in one direction; doing two tasks — vision plus text — either make the model larger or add more data.\"\n\nApple's MM1 technical report pointed at a similar conclusion: the choice of vision encoder, image resolution and the number of vision tokens all affect outcome; higher resolution and more vision tokens usually bring higher training and inference cost. For tasks that only need a few fields read out, having a general model stare at the whole picture is not necessarily cheaper than OCR or a specialist model.\n\n## The timing split: spend compute on multimodality or on Coding?\n\nDeepSeek founder Liang Wenfeng has said: \"To train AI well, you don't need a world model — you don't even need multimodality.\" He also said: \"But you eventually have to do multimodality.\" The two statements do not contradict. Frontier labs do not really disagree on whether multimodal matters long term; the disagreement is on timing and cost.\n\nIn late June, Zhipu chief scientist Tang Jie polled users on X about what features a next-gen GLM \"must have.\" The repeated answer in replies was \"vision.\" Public signals suggest the next-gen base models from Zhipu and Tencent Hunyuan could move further toward native multimodality.\n\nBut DeepSeek V4-Flash has already shown that, without adding vision and without dramatically scaling parameters, post-training can still lift Coding and other core task capabilities substantially. That makes today's resource-allocation question very concrete: spend vision compute and data budget early, or keep pushing scores on Coding and Agent first?\n\nSeveral industry voices argue that, in a phase where Agent task chains keep getting longer, Coding is the price of staying at the table — if Coding does not move up, there will be no future seat. In other words, Coding and Agent capability monetizes over months or weeks; moving a model from \"seeing\" to \"understanding\" is a longer evolutionary cycle.\n\n## Take: two clocks running at once\n\nOn multimodality, Chinese frontier base-model makers are running two different time-races at the same time.\n\nPulling vision into the main model early means \"looking ahead from the present\" — the bet is that, once Agents take over more long-chain tasks and web pages and GUIs become the dominant working surface, visual feedback shifts from a bonus to an entry ticket. That is K3 and Qwen3.8-Max's wager.\n\nSetting vision aside, pushing parameters down and putting every ounce of pressure into post-training means \"holding the line right now\" — the bet is that Coding and Agent leaderboards keep refreshing, and that ranking sits closer to cash flow than \"being able to see\" does. That is DeepSeek V4-Flash's wager.\n\nThese routes are not mutually exclusive. Kimi's strategy has always run post-training in parallel; DeepSeek will do native multimodality eventually. But at this 2026 summer moment, each company has drawn a clear line between training cost and product decisions — and that line, more than any technical verdict, reflects each company's read on tempo. When the two lines finally cross is left for the next delivery cycle to answer.\n\n*Source notes: factual claims in this piece — model parameters, benchmark scores, release timelines, quoted remarks, technical-report conclusions — trace back to the original report [36Kr, \"Kimi K3 与 DeepSeek V4 之间,隔着原生多模态的时间差\"](https:\u002F\u002F36kr.com\u002Fp\u002F3924826666301831?from=newsforai-twoclock) and the Kimi K2.5 technical report.*","native-multimodal-vs-posttraining-2026","2026-08-05T00:00:00Z","2026-08-04T22:11:37.331393Z","2026-08-04T22:11:37.331401Z",true,"agent",128,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"2d29aa3d-317c-4126-a6f7-2c9c2c3b6f93","Kimi K3与DeepSeek V4之间,隔着原生多模态的时间差","kimi-k3-deepseek-v4-native-multimodal-divergence","2026-08-04T08:02:10+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"1844afb1-3a1c-4acd-9e4c-f5e2792a2018","下载免费不等于商用免费：HF Summer 2026 隐藏的开源前沿许可证分水岭","frontier-license-shift-hf-summer-2026","2026-08-23T12:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"4bb31ede-b9c4-4762-86ae-9d3b008557ca","Hugging Face Summer 2026 报告:Qwen 拿下 15 万衍生模型, GGUF 仓库一年涨 464%","hugging-face-state-of-open-models-summer-2026","2026-08-18T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"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":69,"title":70,"news_slug":71,"published_at":72},"40095b51-97b0-4fd4-9b1d-f636c970572e","阿里 Qwen 团队发布 Qwen3.8-Max:2.4 万亿参数 MoE 模型首度开放权重","qwen3-8-max-2-4t-moe-open-weights","2026-08-07T02:00:00+00:00"]