Gemma-4-26B-A4B with 39-69% less thinking at same accuracy
ThinkingCap 发布基于 gemma-4-26B-A4B-it 的优化模型,在 12 项基准上思考 token 平均减少 38.9%、最多减少 69.1%,平均准确率下降 0.15pp。
ThinkingCap 发布基于 gemma-4-26B-A4B-it 的优化模型,在 12 项基准上思考 token 平均减少 38.9%、最多减少 69.1%,平均准确率下降 0.15pp。
Canva 发布首款自研图像生成模型 Canva Craft,所有用户均可通过 Dream Lab 和公开 API 使用。Canva 称,该模型主要基于数百万设计模板从零训练,先规划构图、比例和文字,再生成最高 2K 分辨率、支持超过 30 种宽高比的图像,可渲染 55 种语言的文字,结果可通过 Magic Layers 拆分为可编辑图层。
MiniMax 宣布公开 M3.1 Flash 权重,该模型总参数为 428B,每 token 激活 23B。模型权重可通过 Hugging Face 和 ModelScope 下载。
MiniMax 开源 M3.1 Flash 模型权重,总参数 428B,激活参数 23B。上线 72 小时内,真实用户处理的 token 超过 100 万亿。
Canva released Canva Craft, its first AI image model, now open to all users worldwide. It uses a two-stage design: plan the layout first, then generate the image. The model supports up to 2K resolution and more than 30 aspect ratios. Canva says this architecture improves text legibility and object placement in complex images.
Why it matters: Canva Craft is trained on human-made templates from the platform's content library, covering concrete design scenarios such as social posts, presentations, infographics and menus.
Canva 发布首个图像生成模型 Canva Craft,已向平台所有用户开放,可从 Dream Lab 标签页访问。模型先解析提示词并规划布局,再生成最高 2K、支持 30 种宽高比的图像;Canva 称其支持清晰文字和不改变其他部分的精确编辑。
Recursive 发布首个语言模型 Flint,AI 系统 Eureka Machine 参与了几乎全部开发环节,模型在四个月内完成。
可灵(Kling)宣布 Kling 4.0 已在 OnSolo 上线,支持单次生成 30 秒原生视频。模型支持在快速动作和大运镜下保持稳定,提供立体声、更准确的口型同步及 21:9 宽屏。单个任务可叠加最多 15 个参考图和 10 个关键帧,以保持角色、镜头与剧情一致。
WorldTUI 基于 Qwen2.5-Coder-0.5B 微调,用 360M 参数模型根据按键预测终端屏幕变化,背后没有真实 shell 或文件系统。训练使用 336k 段真实终端会话,并通过错误修复训练改善一致性,通过并行生成屏幕行和顺序验证加速推理。
Recursive 发布可在笔记本电脑上运行的免费编码模型 Flint,称自动化研究系统 Eureka Machine 协助了几乎全部开发环节,小团队在不到 4 个月内从零完成模型。
Inception 发布扩散语言模型 Mercury 3,称其在多项基准上超过 GPT-6 Luna 和 Claude Haiku 5.5,运行速度快 4–8 倍。
NVIDIA 在 GTC Berlin 宣布欧洲车企与出行服务商展示基于 Hyperion 的 L3、L4 及未来 robotaxi 应用,并发布自动驾驶模型与安全技术。
NVIDIA 扩展开放模型组合,推出 Nemotron 3.5 Super VL、Kumo、Cosmos 3.5 等模型,支持组织用本地数据定制并部署到自主控制的基础设施上。
Meta's open-weight model Muse Spark Open went live on Fireworks serverless on launch day, running on US capacity with no US premium. The MoE model has 816B parameters, activates 42B per token, and offers a 1M-token context window for long-horizon coding and agent tasks.
Why it matters: Fireworks runs Muse Spark Open on US capacity with zero data retention by default, giving regulated professional work clear data residency and retention boundaries.
Elevenlabs 发布 Eleven Realtime v1 Max 和 v1 Mini 两款语音到语音模型,面向实时 AI 对话。v1 Max 在 Big Bench Audio 基准上得分为 98.8%,响应时间为 0.57 秒,面向客服等复杂任务;v1 Mini 主打低成本自然对话,能区分用户说完话与仅发出“嗯哼”等回应。
fal 发布与 Black Forest Labs 合作后训练的视频模型 FLUX 3 Max,在第三方独立人类偏好评测中获 1139 Elo,与 Seedance 2.5 并列图生视频质量第一梯队。该模型约 12 秒生成一段 10 秒 720/768p 视频。未来三周降价 75%,后续还将继续通过后训练改进。
Mistral AI 发布面向实时语音智能体的 Voxtral 3 模型家族,将听取、思考、工具调用与说话整合进单一流式管线。系统包含 Voxtral Small Transcribe 3、Voxtral Mini TTS 3 和基于 LiveKit 的语音智能体 SDK,转录覆盖超过 50 种语言,语音输出支持超过 35 种语言及 50 种方言,并支持情绪适配。
Electric Plant 用过去三年收集的数据构建大语言模型 Bombadil,可解读约 125 种植物的电信号并向种植者反馈需求。传感器通过穿透形成层的螺钉采集信号,数据经低功耗无线网络传输后由云端模型处理,结果可用于调整灌溉或施肥安排。
Mistral AI released the Voxtral 3 family of speech recognition and speech generation models, plus a real-time voice agent SDK that combines listening, reasoning, tool calls and speaking. Transcription covers 50+ languages with under 500ms real-time latency. Speech generation covers 35+ languages and 50+ dialects, adjusting emotion, pace and accent to context.
Why it matters: Mistral folds recognition, reasoning, tool calls and speech generation into one streaming pipeline, with each stage swappable so a language or voice change doesn't mean rewriting the agent.
Cohere 发布企业检索重排模型 Rerank 5 Pro 与 Rerank 5 Fast,分别面向重排质量优先和延迟、成本敏感的场景。
Cloudflare 发布文档数据提取模型 Clairic,已在 Workers AI 开放预览,并计划很快开放权重。模型单次处理即可返回用户指定字段的值、格式和置信度,缺失值返回 null,也支持简单计算、估算和预测。
doubleAI 推出代码执行模型 Maxwell,根据程序运行时的硬件与数据优化代码,并对每项改进进行形式化等价证明。它在 CPU 和 GPU 上运行,收集足够运行时数据后,在程序持续运行期间替换为针对当前机器的优化代码。页面称数据处理流水线加速 38 ×、AI 推理内核加速 4.25 ×,LTX-2.5 在两块 B200 上生成每段 4K 视频的耗时从 250s 降至 134 s。
Latent Space 汇总 DeepSeek V4.1 Flash 1017、MiniMax M3.1 Flash 和 Mercury 3 的发布信息、定价与评测,并判断它们暂时不会得到广泛使用。
fal 发布与 bfl_ai 合作的后训练视频模型 FLUX 3 Max,在第三方图生视频质量评测中并列第一。生成 10 秒 720/768p 视频约需 12 秒,未来三周提供 75% 折扣,并被标注为同质量级别最便宜的前沿视频模型。
可灵发布 Kling 4.0,支持单次生成最长 30 秒视频、双声道立体声音频及超过 10 种语言的口型同步,并推出轻量版 Kling 4.0 Flash。主模型支持最多 10 张关键帧图像、8,000 tokens 提示词及最多 15 个参考素材,可通过 MCP 和 CLI 接入自动化工作流;Flash 侧重更快、更低成本的生成,上限为 720p、20 秒。
MiniMax released the reasoning model M3.1 Flash. Its Artificial Analysis Intelligence Index rose to 42 from 29 on the previous M3, while cost per task fell to under half, about $0.22.
Why it matters: M3.1 Flash costs less per task than DeepSeek and leads on hard Q&A and factual accuracy tests, though its overall intelligence index is slightly lower.
Inception released Mercury 3, a diffusion-based large language model. The company says it beats GPT-6 Luna and Claude Haiku 5.5 on several benchmarks and runs 4–8x faster.
Why it matters: Mercury 3 scales past 40x the parameters of Mercury 1 while keeping similar speed, giving a like-for-like read on text diffusion at larger size.
Inception 的扩散模型 Mercury 3 上线 OpenRouter,生成速度超 1,000 tok/s,支持 1M 上下文,为 Mercury 2.5 的 4 倍。
DeepSeek 发布 DeepSeek V4.1 Flash 1017,在 Artificial Analysis Intelligence Index 得分 45,较 DeepSeek V4.1 Flash 提高 6 分。该模型与 GLM-5.3 并列开源权重模型第二,仅次于 MiMo-V2.6-Pro。
Laminar 发布 flow-1,用 RL 训练模型查找智能体轨迹中的错误,并称其在自家基准上检测质量匹配 GPT-6-sol、成本低 23x。模型由小型开源模型微调并经数千次合成调查训练,在 Signals 中检索轨迹片段,按用户定义的 JSON 结构返回发现。
Nvidia 周四发布 Nemotron 3.5 Super VL 和两款轻量 Bolt 开放权重模型,分别面向企业智能体与本地电脑部署。Super VL 拥有 120-billion 参数,支持组织在自有数据上后训练并部署到自有基础设施;Nemotron 3.5 Bolt 74B-A9B 24GB 面向本地多步骤推理、工具调用和网页浏览,Bolt E4B 面向对话智能体。
蚂蚁 inclusionAI 开源 Ming-3.0-tts,提供 0.5b 稠密模型和基于 Ling V3 的 tiny MoE 模型,后者为 7.9B total · 1.3B active。
erikkannike 发布 qwen3-4b-deslop,以 Qwen3-4B-Instruct-2507 为基础在 MacBook Air 上微调,无需系统提示词即可减少破折号、三项列举等 AI 写作习惯。
MiniMax released M3.1 Flash, which it says runs online inference nearly three times faster and cuts overall usage cost on agent tasks by 70% versus the previous generation. The model scores 73.7% on DeepSWE v1.1 and reaches 200+ TPS in everyday output. Pricing is $0.15 per 1M input tokens, $0.60 per 1M output tokens and $0.015 per 1M cached input tokens.
Why it matters: MiniMax claims a 70% cut in agent-task usage cost over the prior generation, with benchmark results on complex tasks, coding and computer use at leading-model levels.
DeepSeek released DeepSeek-V4.1-Flash-1017, which beats its predecessor and some larger competitors on several of DeepSeek's own coding and agent benchmarks. It is available through the website, app and API, and the model weights are open.
Why it matters: DeepSeek reports cutting cost per task from about $0.50 to under $0.20 on its own coding benchmark while scoring higher accuracy than the previous version.
DeepSeek V4.1 Flash 1017 is a model snapshot dated October 17, 2026, served by DeepSeek under the API model name deepseek-flash. It uses a CED architecture and native image understanding, and cuts cache memory to roughly a quarter of the previous generation.
Why it matters: The snapshot cuts cache memory to about a quarter of the previous generation, lowering cache overhead for multi-step agent tasks.
DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017, with the official web and app updated to match. API pricing stays the same and the model is still called through deepseek-flash.
Why it matters: DeepSeek improved the model while holding API price and calling method fixed, and cut task runtime and output length to lower task cost.
Reflection released Beam, an open-source MoE model with 501B total parameters and 23B active per token. It supports a 1M-token context under the Apache 2.0 license.
Why it matters: Hugging Face's guide lays out per-precision deployment and fine-tuning recipes, listing model weights, long-context caching and the hardware each needs.
Meta released Muse Spark Open, an open-weight model with 816B total parameters, 42B active parameters and a 1M-token context window. The weights are available on Hugging Face under the Muse Model License.
Why it matters: Meta ships auto-tunable cross-hardware kernels and post-training support with the model, cutting the cost of building and maintaining custom kernels.
Meta's open-weight model Muse Spark Open launched on Together AI today, aimed at long-running agentic coding workflows. It offers a 1M-token context window and supports sub-agent collaboration, tool calls, computer use and web browsing, and accepts text, image, audio and audio-video input.
Why it matters: Together AI serves Muse Spark Open on launch day and publishes per-item pricing, giving a clear basis for both access and call cost on a long-task model.