SWA bounded replay shows how KV-cache reuse for agent loops can be engineered around sliding-window attention, a pattern many teams will copy as agent traffic grows. Engineers running agent stacks on vLLM can apply it to reduce TTFT under high concurrency. The writeup also signals that DeepSeek-V4.1-Flash inference performance is being actively tuned by the community rather than relying on the lab alone.
Latest
Browser Use 拿 Haiku 5.5 跑了自家基准 v2.1,性能和 Luna 打平但贵 2.8 倍,长任务下还触发 5 倍加价,选型前可以看看这份实测。
生数科技的 Vidu Q4 Preview 一上线就冲到 Artificial Analysis 视频榜第 3,能出 4K 带音频的视频,价格还和上一代一样,可以试试。
Sparse agent memory changes the economics of long-context agent runs, since cache footprint rather than raw context length becomes the cost driver. Developers running multi-step agent pipelines can weigh Flash's cheaper, lighter memory against the extra per-token compute it needs. At US$0.27 versus US$0.67, it also undercuts DeepSeek's own V4-Pro on price while scoring higher on the AA Index.
Anthropic 出了个小杯 Claude Haiku 5.5,价格比上一代还便宜,博主猜参数不到 10B,跑轻量任务可以看看。
Anthropic 的新款小模型 Haiku 5.5 便宜了九成,上下文还扩到 100 万 token,跑高并发任务的成本直接打下来了,做客服、分类、子代理的团队可以看看。
The evaluation suggests routers are judged less on raw cost-performance and more on how quickly they adapt to user feedback, where Jev Router nearly matches Claude Opus 5.5 (High). Teams building multi-model agent pipelines can use these latency and cost figures to decide whether a routing layer beats picking DeepSeek V4.1 Flash directly. The full thread contains additional session-level insights.
284B 的 DeepSeek V4 Flash 量化到 1.6 bit 只要 60GB,配 128GB 内存的 Surface Laptop Ultra 就能本地跑,视频都放出来了。
Claude 出了 Haiku 5.5,跑分压过 GPT-6 Luna,价格一样但速度快两倍,作者说 max 档最香,还预告了测评。
LMArena测了Jev Router的实际延迟:中位数6.18秒,比DeepSeek V4.1 Flash的3.64秒慢了一截,P90差距更明显。
Quantised GGUF builds bring Liquid AI's decision models to laptops and edge devices without full precision hardware, making on-device evaluation feasible for developers. Anyone building scoring, routing or judge pipelines can pick the 1.66 GB d1-3B AD-Q4_K_M quant as a starting point. Downloads are live on Hugging Face, so testing locally requires only a llama.cpp-compatible runtime.
Anthropic 把 Haiku 5.5 做到成本砍掉 75%,电脑操作跑分翻了几倍,跑批量和子智能体任务的成本党可以看看这个。
Anthropic 的新 Haiku 5.5 能手动调 effort,简单任务省钱、难任务加智商,成本你说了算。
Bindu Reddy 直接开炮,说 Haiku 5.5 不如 DeepSeek 和 Luna,Anthropic 该发的其实是 Fable 5.5。
Simon Willison 拿骑自行车的鹈鹕测试了新旧两代 Haiku,5.5 画得比便宜 10 倍的老版本好太多了
Anthropic在Databricks上发布了Claude Haiku 5.5,比前代质量提升15%且成本更低,适合需要高性能小模型的开发者。
Half-price input and free output make this one of the cheapest scoring models for routing, classification and judge-style pipelines, and the open weights allow self-hosting. Teams building agent stacks on OpenRouter can swap it in for v1 without code changes since it is a drop-in version bump. Developers comparing options should check the Hugging Face Decision Index 0.3 leaderboard directly.
The updated benchmarks reveal increasingly subtle differences between top AI music models as quality improves, making specialized evaluation crucial for developers and content creators. For music professionals in ASEAN, these rankings provide concrete guidance on model selection, with Suno v6 currently leading while open-source options like MiniMax Music 3.0 offer accessible alternatives.
有人实测两个同价模型做同一个 3D 场景,一个花 24.96 美元一个只花 1.91 美元,选模型前看看能省不少。
vLLM 花三周把 DeepSeek-V4.1-Flash 跑快了 1.9 倍,AgentX 上吞吐 5.3 倍,还附交互图表拆解过程,做部署的可以看看。
Codex GPT 6 Astra 一次生成就把 Discs of Tron 完整写进 FPGA 逻辑,直接跑在 Chromatic 上,看到演示你会想去试。
Liquid AI 的 d1 3B 单卡跑 Snake,8ms 决策吃了 86 个苹果,OpenAI API 345ms 才吃 3 个,对比很直观。
Matt Shumer 抛了个省钱思路:Opus/Sonnet 当大脑,Haiku 当干活的子智能体,搞智能体的可以试试。
Mistral 的 Mistral Large 4 来了,1T 参数只激活 49B,权重月底开源,视频里直接和 GPT-6.1、Claude Opus 5.5 掰手腕,可以看看它到底排在哪一档。
Anthropic 员工放出的 Haiku 5.5 对比信息:比 Haiku 4.5 快还便宜 75%,编程和 computer use 都升级了,做应用的可以关注价格变化。
一条关于幻觉问题正在被缓解的观点帖,作者认为半年来的模型进步改变了局面,可以点链接看看他给的论据。
Anthropic 说 Haiku 5.5 在 10 万 token 以下的任务最划算,而这占了日常请求的九成,用小模型的可以算笔账。
Claude Haiku 5.5 已经能在 OpenRouter 上用了,比上代便宜 75%,还支持推理力度,速度实测 100+ TPS,快速场景可以试试。
OpenAI 把 GPT-6 直接推给所有 ChatGPT 用户了,还带智能界面,回答能可视化,复杂问题自动生成可操作的小工具,值得上手试试。
Anthropic 的小模型 Haiku 5.5 便宜了 90%,操作电脑的基准分从 15.7% 拉到 72.4%,跑批量智能体任务的可以看看。
OpenAI 把新版 GPT-6 推给全部 ChatGPT 用户了,这次不只是文本,还做了不少底层基建改造,撑得起 12 亿用户。
Anthropic 新出的 Haiku 5.5,100k 上下文,比 Haiku 4.5 便宜 10 倍,跑批量任务的成本压力小多了。
Anthropic 放出了 Claude Haiku 5.5,比 4.5 便宜七成五,还能在 Agent Arena 里跑真实任务,去试试手感。
Anthropic 的 Haiku 5.5 来了,智力比一年前的 Haiku 涨了 26 分,价格还只有上代的十分之一,跑 agent 任务可以重点看看。
Anthropic 又发新模型了,Haiku 5.5 比 4.5 便宜 75%,跑批量任务可以直接用它,重的活留给 Opus。
Only the latest 2000 items are listed; for older ones see the daily archive