The notable detail is cost efficiency: Hy4 preview nearly matches Kimi K3 (Max) on agent performance while cutting per-task spend by about two-thirds, placing it on the Agent Arena Pareto frontier. Teams running open-source agent stacks on a budget should benchmark it against Kimi K3 (Max) when the full Hy4 release lands. Its signal-level strengths in Bash Recovery and tool reliability matter for CLI-heavy workflows.
Particle-1 上线 OpenClaw 了，输出 token 比 DeepSeek V4.1 Flash 少 45%，跑 agent 的成本能省不少，升级到 v2026.10.10 就能用。
Abacus.AI 把音频决策模型 Smaug-Audio-Decisions 开源了，Hugging Face 就能下，做语音 Agent 的可以拿来试试。
A personality rewrite this large between consecutive versions means teams tuning prompts or evals around Kimi K3's behavioural profile may see different outputs after upgrading. Developers building character-driven agents or customer-facing assistants should re-run behavioural evaluations before switching to K3.1 Pro. The benchmark also offers a reusable method for tracking such shifts across any model lineage.
Sage2 把自然语言直接转成结构化数据，还能边接收上下文边更新判断，官方称成本比 Opus 5.5 低 166 倍，做实时客服可以看看。
BAAI 和北邮、北大搞的 LLaCO，8B 连续扩散模型，数学代码上干过了 LLaDA-8B 和 Dream-7B，想追扩散语言模型路线的可以看看。
Fewer output tokens at similar accuracy directly cuts API costs, since token count drives latency and spend on agentic workloads. Zero data retention makes the model a candidate for handling sensitive prompts. Builders using OpenClaw can try it by updating to v2026.10.10, and teams should compare its cost per task against DeepSeek V4.1 Flash on their own workloads.
腾讯混元这个 Hyra 2.0 让 AI 自己设计模型、写代码、做训练，从蛋白质工具到修香山芯片都干过了，看看三层智能体怎么分工的。
微信视觉团队开源了 WeVisOne-VL，一个模型搞定检测、计数、OCR 和抠图，才 9.6M 参数解码器，不用再拼一堆工具了。
GPT-6 Astra 只改了 Qwen3.5-0.8B 万分之十六的参数，没重训就把数学成绩从 7.9% 拉到 48.1%，挺离谱的实验。
有人发现 Astra 已经在偷偷路由到 GPT-6.2 了，还能直接生成 SVG，看来 GPT-6.2 可能马上要正式发布。
Attention variants like MLA, Mamba-2 hybrids and sparse attention are replacing vanilla transformers as the default choices in frontier open weights, so this roundup is a fast map of which lab bet on what. Practitioners choosing models to self-host can scan parameter counts and mechanisms before committing GPU budget. Since most of these weights are openly released, teams in Singapore and ASEAN can evaluate them locally without waiting for regional API availability.
OpenBMB 放了个演示，MiniCPM-V 4.7 在手机本地就能操作真实 App 完成多步任务，不联网不靠云，端侧 Agent 这块可以看看。
One small model replacing a stack of separate detectors, OCR engines and matting tools simplifies deployment pipelines for practitioners building document or retail vision systems. The model and data pipeline are on GitHub at github.com/Tencent/WeVisOne-VL, so teams can self-host the 2B or 4B weights. Compared with Qwen3-VL or dedicated segmenters, it trades a single specialised leaderboard edge for broad task coverage plus pixel-level outputs at low parameter counts.
The pricing detail matters more than the demo: cached-input economics at $0.006 per 1M tokens can make long-horizon agent runs dramatically cheaper than headline per-token rates suggest. Developers building multi-step coding agents can compare LongCat-2.5's cache-heavy cost profile against Qwen3.8-Max on their own workloads, with both models available via longcat.ai. The 1M-token training sequences also point to better endurance on tasks that span dozens of agent turns.
The cross-generation sweep is the real signal: weight-editing ability appeared abruptly in the latest frontier models, echoing the stage-like jumps seen in interpretability research. Practitioners tracking emergent capabilities or open-source progress should watch the planned replication on open weights, since local models eventually editing their own weights is the stated end goal. Compared with prior manual work on networks of tens to hundreds of nodes, this extends the reachable scale, though the task itself remains simple and small.
有人拿 Opus 5.5 写代码复刻了 OpenAI 宣传片，全程 Three.js 加 GLSL，一个视频生成模型都没用，可以看看效果。
Mistral 把 ML4 的权重直接开放了，代码智能体和网络安全能力是主打，三周内比 Preview 版瘦身不少，技术报告也一起放出来了。
有人晒出 Fable 5.5 纯代码生成动画的视频，还附了验证方法，你可以自己跑一遍提示词对比看看真假。
有人用 Fable 5.5 一句话生成了纯代码动画视频，同提示词在 Opus 5.5 和 Fable 5.1 上效果都明显更差，可以照着截图里的提示词自己试。
Shikhar 用一次请求就生成了 30 秒蜘蛛侠风格的 WebGL 动画，全塞进一个 HTML 文件，还有人发现 Opus 请求被路由到了 Fable 5.5。
网友爆料 Anthropic 有个叫 Fable 5.5 的 Claude 检查点流出，据说纯写代码就能做 3D 动画，不用任何视频模型，不过官方还没承认，先当八卦看。
Particle 的模型在 Plotly 数据分析基准上超过了 Terra-6.1 和 Haiku 5.5，是新玩家多年首次上榜，做数据分析的可以看看评测细节。
Krea 的 K3 Edit 图像编辑模型已经能在 OpenRouter 直接调用了，想试图像编辑的可以去玩玩，不用单独注册 Krea 账号。
Camera-controllable world models turn static images into navigable scenes, which matters for robotics simulation, virtual production and synthetic data generation. Builders working with NVIDIA's Cosmos stack can read the post-training recipe directly and replicate the camera-control fine-tuning. The report offers a concrete reference point against other controllable video generation approaches like camera-path conditioned diffusion models.
Levanto Labs 发了个抽取模型 Sage2，说字段就返回带原文出处和概率的结构化数据，比 Cloudflare 的 Clairic 还强，便宜 Opus 5.5 166 倍，做论文信息抽取可以试试。
A single-prompt run producing a working scientific solver suggests frontier models are moving past code snippets toward end-to-end technical workflows with self-validation. Researchers and engineers in computational chemistry or scientific computing can test whether the same prompting pattern transfers to their own simulation tasks. It also gives a concrete point of comparison against Claude, GPT-5 and Gemini on domain-specific, multi-step reasoning.
Fable 5.5 不用图像模型，纯靠写 Python 代码逐帧画出一部 66 秒动画，连配乐音效都是代码生成的，这条演示值得看看效果。
LMArena 上线了智能体排行榜，想给 Agent 任务挑模型的话可以直接去 arena.ai/leaderboard/agents 按类别对比各家成绩。
A specialised open model beating a frontier closed model on a science benchmark strengthens the case for domain-specific post-training over general-purpose scale. Research labs and computational chemistry teams can download and adapt the weights for their own discovery workflows. Its base model, Zhipu's GLM-5.3, also shows Chinese open models being repurposed at the frontier of applied science.
SemiAnalysis 搞了个新基准 GDPx，最强模型只拿 12%，里面 Gemini Carbon 在智能体任务上表现挺亮眼，可以看看。
有人做了个 DeepSeek 相关的原型实验，注意是合成负载、部分实现，官方还没出对应产品，看解读前先知道这个前提。
DeepSeek 公开的多 Agent 架构，实验数据很扎实：单会话 p95 要 29 分钟，Super-agent 15.7 分钟搞定，代价是 token 翻几倍。
SemiAnalysis 弄了个智能体新基准 GDPx，最强模型才拿 12% 分，Pichai 还点名 Carbon 的成本表现，想看模型真实水平差距的可以瞅瞅。
SemiAnalysis 搞了个新基准 GDPx，最强模型才拿 12 分，想看各家智能体真实水平差距的可以瞅瞅这份榜单。
DeepSeek 的 V4.1 Flash 1017 把开源模型在 ARC-AGI-2 拉到 78.1%，还顺手刷了 ARC-AGI-3 和 -1 的开源最佳，关注开源推理的可以看看成绩单。
Only the latest 2000 items are listed; for older ones see the daily archive
