Latest
微软基于 Qwen3.5-9B 出了个决策小模型，专管 Agent 流程里的判断打分，比 GPT-6 Sol 快 35 倍，输出还免费。
微软把决策类小模型做成了专项产品，比 GPT-6 Sol 快 35 倍还输出免费，做分类和判断任务的可以试试。
Causal tracing over real long-horizon sessions could surface failure modes that preference voting hides, since voters rarely see multi-step trajectories end to end. Teams evaluating agents for production workflows should check the five signals to see which align with their own quality metrics, and compare scores against Chatbot Arena-style leaderboards before picking a model.
This specialized decision model addresses a critical inefficiency in AI applications - replacing expensive general-purpose models for simple choice-based tasks. For developers in Southeast Asia, it offers significant cost reductions (up to 200x) and speed improvements (up to 100x) for tasks like content moderation, intent recognition, and AI output evaluation. The model's availability through Microsoft Foundry and OpenRouter makes it immediately accessible for production use cases across the region.
Bindu Reddy 爆料 Google 内部已有比 Gemini Pro Argon 更强的模型，Argon 会不会发布即落后？
微软新发布的决策模型Microsoft-Decision-1，在结构化决策任务中表现优异，已应用于科学发现等场景。
微软用 Qwen3.5-9B 做了个专门快速打分决策的模型，比 GPT-6 Sol 快 35 倍，Foundry 上就能用。
The interesting question now is serving: distilled few-step models like this shift the bottleneck from sampling to the inference stack, which is why vllm-omni's acceleration work matters for real deployments. Developers running Qwen-Image in production should benchmark this turbo variant against their current full-step setup to see whether the 8-step mode lets them cut GPU cost per image. It also gives open-source users a faster default that reduces the gap with proprietary image APIs.
微软出了个专做决策的小模型，不走聊天路线，主打低延迟，公司内部拿来做事故响应和质量控制了，可以看看定位。
StepFun 的 Step 5 Preview 限时免费，600B 参数还能看图，得分和 GPT-6 Luna 打平，可以先去试试再决定要不要付费。
The takeaway is that single-edit benchmarks can mislead: models that re-render the whole image, like GPT Image 2.5 (Sunburst), accumulate drift across repeated edits, while local editors like Ideogram 4.5 and FLUX 3 preserve detail. Teams building multi-turn design or staging workflows should pick models by edit locality, not leaderboard rank. Real estate staging, product photo iteration and e-commerce pipelines in any market can apply this 30-step test method to their own candidate models.
A 1M-token context with self-checking behaviour makes this a practical pick for unattended long agent runs on unfamiliar codebases. Developers can plug it into existing tools like Cline or OpenCode without changing workflows, and the open-weights release on 15 Oct lets teams self-host. Its strong frontend generation and document-to-deck abilities also put it in direct comparison with frontier coding models for everyday engineering work.
有人爆料 OpenAI 秘密模型 Aeon，一道题平均想 3 小时就啃下不少数学难题，训练曲线还没到头，看着挺吓人。
Leaderboard rank alone overstates the win: the cheaper Alibaba Qwen model actually beats HeyGen on pronunciation robustness, and Eleven v4 is nearly twice as fast, so the top spot rests on voice quality in assistant-style use cases. Teams building voice agents in US and UK English can use the Elo splits by category, plus the price and speed data, to pick a TTS vendor. Alibaba's strong showing also signals competitive Chinese TTS at lower cost.
The 45% figure reframes the debate: the gap is not raw coding ability but architectural judgment, prod-risk reasoning and knowing what not to build. Engineering leaders can use sudo L7's rubric categories to decide which parts of a staff engineer's job are safe to delegate to agents today. Teams evaluating coding agents against internal tickets should note the benchmark draws on private production repos, a harder and more realistic setting than public-repo suites like SWE-bench.
四个模型改同一张图 30 次，Ideogram 4.5 几乎不动画面，GPT Image 2.5 越改越走样，做图的朋友值得看看测试方法。
Nvidia 搞了个 ARC，不收集新机器人数据、不改架构，就能让机器人模型在推理任务上成功率翻 5 倍以上。
Nano Banana 2.1在图像生成领域表现亮眼，Mistral Large 4在欧洲实验室中排名领先，Claude Haiku 5.5性价比突出。
OpenAI 称 LLM 在 7 个千禧年难题中的 4 个上有实质进展，连 Claude Opus 都来背书，但结果还得通过验证，这条帮你理清边界。
Mistral 上了个 1T 参数的大家伙，激活只有 49B，Agent Arena 排名升了 11 位，10 月底开源，可以等等看。
35B 的模型离线跑在 iPhone 上只占 1.8GB 内存，这压缩水平真少见，做端侧 AI 的朋友可以看看他们的开源方案。
一个不走 VAE 压缩、直接逐像素画图的开源模型，3B+4B 的组合不大，效果一般但架构挺新鲜，搞生成的可以看看。
Anthropic 刚发了小模型 Claude Haiku 5.5，成本比 Haiku 4.5 便宜约 75%，Genspark 的聊天和代码功能已经能直接用了。
ExplorationBench addresses a critical gap in AI evaluation by focusing on discovery rather than recall, revealing that even advanced systems struggle with rule application despite knowing them. This matters for researchers developing scientific AI tools and for organizations deploying AI in novel domains, as it highlights the importance of feedback mechanisms and experimental design. The benchmark is now available for evaluation at explorationbench.com, with code coming soon to GitHub.
Only the latest 2000 items are listed; for older ones see the daily archive