Device storage is unavailable. Reading preferences will last for this visit only.
EXPLORE A TOPIC
AI model releases, research, benchmark results, and version changes.人工智能模型发布、研究结果、基准测试与版本变化。
Canva launched Craft on October 22 and says its new research lab built the model around design layouts, text, and editability. The release also gives a rare look at the lab’s size, while its strongest rankings and quality claims still come from Canva’s own evaluation.Canva 于 10 月 22 日推出 Craft，并表示新成立的研究团队围绕设计布局、文字和可编辑性构建了这款模型。此次发布也首次披露了研究实验室的规模，但最强的排名和质量主张仍来自 Canva 自己的评估。
GitHub confirmed the retirement of six model selections across most Copilot experiences. The final notice names replacements, but availability still depends on plan, client, and administrator policy, so a model change should be tested in the surface where a team actually works.GitHub 确认在大多数 Copilot 使用场景中弃用了六个模型选项。最终公告列出了替代模型，但用户仍需检查套餐、客户端和管理员策略，团队应在实际使用的界面中测试迁移结果。
Reflection published Beam 1.0’s 501-billion-parameter weights in three precision formats on October 16. The Apache 2.0 license gives teams room to inspect and adapt the model, but a very large checkpoint still needs serious memory, serving infrastructure, evaluation, and guardrails.Reflection 于 10 月 16 日发布了 Beam 1.0 的 5010 亿参数权重，提供三种精度格式。Apache 2.0 让团队可以更自由地检查和改造模型，但庞大的检查点仍需要充足内存、服务基础设施、评测和防护措施。
NVIDIA released a 120B-parameter model that accepts text, images, and video for workflows involving documents and tools. Its headline speed claims describe specific vendor tests, while an independent benchmark measures a different slice of the model’s performance.NVIDIA 发布了一款模型，能接收文本、图片和视频，用于文档与工具工作流程。其醒目速度主张来自特定厂商测试；独立基准则测量了另一部分性能。
Meta released the model weights on October 15 and showed them running across linked Mac Studios. That demonstration is one way to host an 816-billion-parameter model; it does not make local operation a one-computer download or remove the license conditions.Meta 于 10 月 15 日发布模型权重，并展示了跨多台相连 Mac Studio 运行的方式。这是运行一款 8160 亿参数模型的一种本地方案，并不意味着普通电脑下个文件就能运行，也不代表许可没有条件。
Meta introduced Wonder and Wonder Full on October 12, then updated its developer documentation with clearer routes and plan limits. The distinction that matters is which model a person can actually access, through which product, and under what data-use setting.Meta 于 10 月 12 日推出 Wonder 和 Wonder Full，之后更新了开发者文档，写明模型路径与计划限制。真正重要的是：用户实际上能通过哪款产品访问哪个模型，以及数据使用设置是什么。
Mistral’s October 6 announcement made Large 4 available through a hosted preview API and promised downloadable weights later in the month. As of October 24, the company’s current changelog still says those weights are coming soon.Mistral 在 10 月 6 日宣布通过托管预览 API 开放 Large 4，并承诺当月晚些时候发布权重。截至 10 月 24 日，公司当前更新日志仍写着权重即将推出。
A Google DeepMind team and University of Maryland professor Matt Landreman published two new families of exact three-dimensional magnetic-equilibrium solutions that break stellarator symmetry. Their October preprint says Gemini agents proposed the generalizations; the authors independently re-derived and checked every derivation.Google DeepMind 团队与马里兰大学教授 Matt Landreman 于 10 月 5 日发布两组新的三维精确磁平衡解，打破了恒星器对称性。该预印本称 Gemini 智能体提出了推广方案；作者随后自行重新推导并用符号计算检查了每一步。
Anthropic positions Sonnet 5.5 for coding, documents, and other well-scoped work. Its per-token price matches Sonnet 5, while the company says it uses fewer tokens per task and adds cyber safeguards.Anthropic 将 Sonnet 5.5 定位为适合编码、文档和其他边界清晰工作的模型。它的每 token 价格与 Sonnet 5 相同，公司称模型完成单项任务时能用更少 token，并新增网络安全防护。