[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-grok-imagine-1-5-lite-cheaper-than-veo":3,"topics-all":38,"news-related-240d8a4c-c025-4080-be14-06099f650a71":57},{"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},"240d8a4c-c025-4080-be14-06099f650a71","xAI 把视频价格打到底:1080p 蒸馏版 Lite 比 Veo 3.1 便宜 65%","xAI 发布 Grok Imagine Video 1.5 Lite,这是旗舰视频模型的蒸馏版。1080p 定价 0.14 美元\u002F秒,比 Veo 3.1 含音频版便宜 65%,在 AA 视频榜排名第 17,比 Veo 3.1 高两位。","## 蒸馏换价格,1080p 单秒 0.14 美元\n\nxAI 这次没有发新架构,而是把现役旗舰 Grok Imagine Video 1.5 蒸馏成了 Lite 版。蒸馏(student model 学大模型行为)的代价是质量略降,但换来的是更便宜和更快。OpenRouter 模型卡写得直接:Lite 是一个\"在速度与价格上做了取舍,1080p 实际由 720p 渲染再升频\"的视频生成模型。\n\n价格表非常激进:480p 0.02 美元\u002F秒、720p 0.03 美元\u002F秒、1080p 0.14 美元\u002F秒。同档 1080p 摆到一起:Veo 3.1 含音频版 0.40 美元\u002F秒,Lite 比 Veo 3.1 便宜 65%,也比完整版 1.5 的 0.25 美元\u002F秒便宜 44%。60 秒总素材成本:Lite 1080p 是 8.40 美元,完整版 1.5 是 15 美元,Veo 3.1 是 24 美元。\n\n## 比 Veo 3.1 高两个名次,落在两条 Pareto 前沿上\n\nArtificial Analysis 在 AA-Video-T2V v2.0 与 AA-Video-T2V-Silent v2.0 两张榜上把 Lite 排在第 17 位,刚好高出 Veo 3.1 两位。AlphaSignal 的报道进一步指出,Lite 同时落在质量-速度前沿和质量-价格前沿,也就是说在它所在的层级里,目前没有其它模型能同时压更低延迟和更低价格。\n\n中位生成延迟 60.5 秒(产出 10 秒 1080p 素材),这个量级足以支撑迭代式 prompt 调试与故事板工作流。横向参考:Kling 3.0 1080p Pro 质量略高一档但 94 秒才能渲染 5 秒片段;Vidu Q3 Turbo 的 5 秒 720p 配置快约 9 秒,但质量评分明显更低。\n\n## 强项在多镜头叙事,短板在唇同步与人体\n\n按 AA-Video-T2V 的十项能力细分,Lite 最接近前沿的是多镜头与叙事(Multi-Scene & Narrative)、光照与材质(Lighting & Materials)、文字渲染(Text Rendering),与完整版 1.5 之间的差距也最小。\n\n短板集中在对话与唇同步(Dialogue & Lip Sync)以及人体解剖(Human Anatomy)。用例层面,它在建筑与房地产、消费内容、生产力与知识工作(产品走查、室内漫游、虚拟陈设、社交短视频、b-roll、解释视频、教育素材)这几类里表现最好。Live-Action Film 和榜单的 Frontier 类目是它的明显短板——对话密集、对人脸\u002F手部要求高的镜头,生产用之前得做针对性 prompt 评测。\n\n## 1080p 是 720p 升频,选型时要看清\n\nOpenRouter 模型卡特别强调:Lite 的 1080p 输出其实是 720p 渲染后再升频,帧尺寸 1080p 但原生细节来自 720p。细纹理、小字、裁切和合成场景,这层差异会显形。第三方目录有的只暴露 480p\u002F720p 两档,依赖 1080p 细节的应用需要把 Lite 的升频输出与完整版 1.5 在自家 prompt 集上做对比。\n\n## 实用工作流建议\n\nLite 不是用来替代 1.5 的,而是用来\"做草稿\"。10 条 1080p Lite 草稿成本 14 美元,挑一条用完整版 1.5 重生成一次 2.50 美元,合计 16.50 美元;直接用完整版 1.5 跑 10 条是 25 美元。把草稿档切到 720p,10 条草稿 3 美元,加一条完整版 1080p 2.50 美元,合计 5.50 美元。\n\nLite 的基准画像最贴合建筑漫游、解释视频、社交短视频、b-roll 和多镜头草稿,需要同步对话、对人体敏感的特写、要求高的实拍场景时,完整版 1.5 或更强专用模型仍然应该顶上。\n\n## Veo 3.1 是被压着打的对手\n\n数字摆出来很直白:同档 1080p,Lite 比 Veo 3.1 含音频便宜 65%,榜单名次比 Veo 3.1 高两位,中位渲染延迟 60.5 秒。视频生成市场的中价位段——基础工具与前沿旗舰之间的缝隙——正在被 xAI 用 Lite 这一档压下来。\n\nVeo 3.1 走\"含音频、长上下文\"路线,Veo 3.1 Fast 走更便宜的另一档;xAI 的 Lite 则直接压在 Veo 3.1 名次之上,价格却只有它的 35%。AA-Video-T2V v2.0 的排名本身依赖一组特定 prompt、评分方法和模型快照,生产环境的真正结论还得叠上各路由可用性、失败作业处理、排队延迟、音频质量以及自家 prompt 集上的实测。Lite 已经把\"同价位最便宜 + 同价位最快\"这件事做出来了,接下来要看的是生态侧——fal、Vercel AI Gateway 这些分发渠道能不能把它快速带到开发者的实际工作流里。","https:\u002F\u002Falphasignal.ai\u002Fnews\u002Fxai-s-grok-imagine-video-1-5-lite-beats-veo-3-1-at-65-lower-cost","b82e17a3-1dbd-4b5d-88dc-9f518f917cc0",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ff27d4a9-31ed-4aa0-ad4e-ab298ba067d7","en","xAI's Grok Imagine Video 1.5 Lite undercuts Veo 3.1 by 65%","xAI released Grok Imagine Video 1.5 Lite, distilled from 1.5. USD 0.14\u002Fsec at 1080p, 65% cheaper than Veo 3.1 with audio. AA-Video-T2V rank No. 17.","## Distillation for price: USD 0.14\u002Fsec at 1080p\n\nxAI did not ship a new architecture here. Instead, the company distilled its flagship Grok Imagine Video 1.5 into a Lite variant, trading some quality for lower cost and faster turnaround. The OpenRouter model card is explicit: Lite trades quality for speed and price, and 1080p output is rendered at 720p and then upscaled.\n\nThe pricing is aggressive: USD 0.02\u002Fsec at 480p, USD 0.03\u002Fsec at 720p, USD 0.14\u002Fsec at 1080p. At the same 1080p tier, Veo 3.1 with audio costs USD 0.40\u002Fsec, so Lite is roughly 65% cheaper than Veo 3.1 and 44% cheaper than the full Grok Imagine Video 1.5 at USD 0.25\u002Fsec. For 60 seconds of total footage, Lite at 1080p is USD 8.40, full 1.5 is USD 15.00, and Veo 3.1 is USD 24.00.\n\n## Two ranks above Veo 3.1, on two Pareto frontiers\n\nArtificial Analysis places Lite at No. 17 on both the AA-Video-T2V v2.0 and AA-Video-T2V-Silent v2.0 leaderboards, two places above Veo 3.1. AlphaSignal reporting notes that Lite also lands on the Pareto frontiers for quality-vs-speed and quality-vs-price, meaning that within its band, no measured alternative improves both dimensions at once.\n\nMedian generation latency is 60.5 seconds for a 10-second 1080p clip, fast enough to support iterative prompt work and storyboarding without long batch queues. For reference, Kling 3.0 1080p Pro scores slightly higher on quality but takes 94 seconds to produce a 5-second clip; Vidu Q3 Turbo 5-second 720p configuration finishes about nine seconds sooner than Lite 10-second 1080p output but receives a substantially lower quality score.\n\n## Strong in multi-scene narrative, weak in lip-sync and anatomy\n\nAcross the AA-Video-T2V v2.0 capability groups, Lite comes closest to the frontier in Multi-Scene & Narrative, Lighting & Materials, and Text Rendering. These are also the dimensions where the gap to the full 1.5 model is smallest.\n\nThe largest deficits appear in Dialogue & Lip Sync and Human Anatomy. On the use-case axis, Lite performs best on Architecture & Real Estate, Consumer Content, and Productivity & Knowledge Work — exterior fly-throughs, interior walkthroughs, virtual staging, social clips, b-roll, explainers, and educational clips. Live-Action Film and the Frontier category show the largest gaps, so dialogue-heavy scenes and anatomy-sensitive close-ups warrant prompt-specific evaluation before production use.\n\n## 1080p is upscaled — read the fine print\n\nThe OpenRouter model card is explicit that Lite 1080p output is rendered at 720p and upscaled: the frame dimensions are 1080p, but native detail comes from a 720p render. For fine textures, small text, cropping, and compositing, that distinction matters. Some third-party listings expose only 480p and 720p for this model ID, so applications that depend on fine 1080p detail should compare Lite upscaled output against the full model output on representative prompts.\n\n## A practical two-tier pipeline\n\nLite is built for drafts, not replacement. Ten 10-second Lite drafts at 1080p cost USD 14.00; regenerating one selected prompt with the full model adds USD 2.50, for a total of USD 16.50. Ten full-model drafts would cost USD 25.00. Dropping the draft tier to 720p brings the bill to USD 3.00 for ten drafts, plus USD 2.50 for one full-model 1080p generation.\n\nThis works because Lite benchmark profile aligns most closely with architectural walkthroughs, explainers, social clips, b-roll, and multi-scene drafts. For synchronized dialogue, anatomy-sensitive close-ups, and demanding live-action sequences, the full 1.5 or a stronger specialist remains the right budget allocation.\n\n## Veo 3.1 is the model getting squeezed\n\nThe numbers line up: at the same 1080p tier, Lite is roughly 65% cheaper than Veo 3.1 with audio, ranks two places above Veo 3.1 on the AA leaderboard, and renders a 10-second 1080p clip in a 60.5-second median. The mid-band of the video generation market — between basic tools and frontier-priced generators — is exactly where xAI is now applying pressure.\n\nVeo 3.1 has been positioned around audio integration and long-context generation, while Veo 3.1 Fast occupies a cheaper adjacent tier. Lite lands on the Pareto frontier above Veo 3.1 while costing roughly 35% as much. AA-Video-T2V v2.0 rankings depend on a specific prompt set, scoring methodology, and model snapshot, so production conclusions should add route availability, failed-job behavior, queue latency, audio quality, and benchmarks against user own prompts. Lite has done its part: cheapest at its tier, fastest at its tier. The next variable is distribution — whether fal and Vercel AI Gateway can move it from announcement to developer workflows quickly enough.","grok-imagine-1-5-lite-cheaper-than-veo","2026-10-10T08:00:00Z","2026-10-10T05:06:54.516881Z","2026-10-10T05:06:54.516890Z",true,"agent",60,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"5757433d-e8ca-4371-89bc-f563dacc8089","程序能跑≠画对了:MaLiang框架专治AI作画跑偏","maliang-harness-p2v-gap","2026-09-30T15:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"ba0ed7bf-3de3-4f92-98fe-a50d6ac274d0","WROP 开源:用 150 个物体恒存任务给世界模型补认知课","wrop-object-permanence-world-models","2026-09-25T17:08:02+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"371dd90f-e8bb-46a1-967a-346c28b393df","MiniMax-H3 会不会推理物理世界?517 题实测只对 42%","minimax-h3-physical-reasoning-eval","2026-09-19T17:11:04+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"6062d551-9068-4a9f-ae8e-4e99269cd838","Muse Voice Transcribe 发布:流式转写、20+ 说话人分离、端点检测,Meta 全塞进一个模型","meta-muse-voice-transcribe-streaming-asr","2026-09-05T13:11:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"7ddc323f-fc52-406a-b6df-79b7393e121b","高德开源 DreamX-Creator:7B 原生音视频生成,2K 输出","dreamx-creator-7b-native-audio-video","2026-09-01T13:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"1bf979b5-2b52-4a19-8e2d-e0a549bb24ba","fal 后训练版 MiniMax H3:5 秒视频约 3 秒生成,吞吐 35 倍","fal-h3-max-post-trained-video","2026-08-29T23:02:12+00:00"]